<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Satyam Kumar</title>
    <description>The latest articles on DEV Community by Satyam Kumar (@satyam_kumar_1229ce0b7f2c).</description>
    <link>https://dev.to/satyam_kumar_1229ce0b7f2c</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4140818%2F5efedfb3-9a4f-4361-865c-0c10d88b30b3.png</url>
      <title>DEV Community: Satyam Kumar</title>
      <link>https://dev.to/satyam_kumar_1229ce0b7f2c</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/satyam_kumar_1229ce0b7f2c"/>
    <language>en</language>
    <item>
      <title>ExperimentPilot AI: Stop Wasting Hours on EDA and Debugging ML Models</title>
      <dc:creator>Satyam Kumar</dc:creator>
      <pubDate>Sun, 18 Oct 2026 07:00:16 +0000</pubDate>
      <link>https://dev.to/satyam_kumar_1229ce0b7f2c/experimentpilot-ai-stop-wasting-hours-on-eda-and-debugging-ml-models-2del</link>
      <guid>https://dev.to/satyam_kumar_1229ce0b7f2c/experimentpilot-ai-stop-wasting-hours-on-eda-and-debugging-ml-models-2del</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week2-2026-10-12"&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  ExperimentPilot AI: On-Device Dataset Behavior Discovery &amp;amp; ML Experiment Analysis Platform
&lt;/h1&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;"Upload your data. Discover how it behaves. Understand why the patterns matter. Decide what to investigate next."&lt;/strong&gt;&lt;br&gt;&lt;br&gt;
&lt;em&gt;Transforming raw tabular datasets and ML experiment predictions into empirical discoveries, statistical evidence, and AI-grounded hypotheses — 100% locally and privately.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. The Core Problem: The Silent Bottleneck in Data Science &amp;amp; ML
&lt;/h3&gt;

&lt;p&gt;Every data scientist, machine learning engineer, and quantitative analyst faces the exact same time-consuming ritual whenever they receive a new dataset or evaluate a new model iteration:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Repetitive, Manual Exploratory Data Analysis (EDA):&lt;/strong&gt; Spending hours writing boilerplate Pandas, Seaborn, and Scikit-learn scripts to check missing values, calculate skewness, generate correlation matrices, run ANOVA tests, and plot PCA clusters.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The "Why Did My Model Fail?" Black Box:&lt;/strong&gt; Standard experiment trackers (like MLflow or Weights &amp;amp; Biases) excel at logging loss curves and hyperparameters, but they leave engineers stranded when answering: &lt;em&gt;Why did Model B fail on samples Model A got right? What specific feature subgroups caused the failure?&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The Cloud AI Dilemma (Privacy &amp;amp; Hallucinations):&lt;/strong&gt; Uploading proprietary enterprise data, healthcare records, or financial figures to closed cloud APIs (OpenAI, Anthropic) directly violates corporate data governance, HIPAA, and GDPR. Furthermore, asking generic cloud LLMs to analyze data leads to fabricated numbers and non-reproducible statistical claims.&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  2. The Solution: ExperimentPilot AI v2.0
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;ExperimentPilot AI&lt;/strong&gt; is an on-device, privacy-preserving &lt;strong&gt;Dataset Behavior Discovery and ML Model Experiment Workstation&lt;/strong&gt;. It pairs &lt;strong&gt;deterministic mathematical computation&lt;/strong&gt; (Pandas, NumPy, SciPy, Scikit-learn) with &lt;strong&gt;local open-weight AI inference&lt;/strong&gt; (via Ollama) to autonomously profile data, discover hidden behavioral patterns, isolate model disagreements, and generate actionable investigation hypotheses.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;┌────────────────────────────────────────────────────────────────────────────────────────┐
│                                 EXPERIMENTPILOT AI WORKFLOW                            │
│                                                                                        │
│   [Raw Tabular CSV] ──► [Deterministic Math Engine] ──► [Compact Evidence JSON]       │
│                               (SciPy / Scikit-learn)               │                   │
│                                                                    ▼                   │
│   [Interactive Q&amp;amp;A &amp;amp; Reports] ◄── [React Dashboard] ◄── [Local Open-Weight LLM]       │
│        (Recharts Visuals)           (100% On-Device)        (Ollama: Qwen 2.5 0.5B)    │
└────────────────────────────────────────────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;The platform delivers two interconnected operational engines:&lt;/p&gt;
&lt;h4&gt;
  
  
  Engine A: Automated Dataset Behavior Discovery
&lt;/h4&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Fine-Grained Semantic Type Inference:&lt;/strong&gt; Automatically categorizes columns into &lt;code&gt;numeric_continuous&lt;/code&gt;, &lt;code&gt;numeric_discrete&lt;/code&gt;, &lt;code&gt;categorical_nominal&lt;/code&gt;, &lt;code&gt;categorical_ordinal&lt;/code&gt;, &lt;code&gt;datetime&lt;/code&gt;, &lt;code&gt;identifier&lt;/code&gt;, &lt;code&gt;free_text&lt;/code&gt;, and &lt;code&gt;constant&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;0–100 Data Quality Health Score:&lt;/strong&gt; Evaluates missingness penalties, duplicate row rates, constant features, and extreme distribution skews.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Automated Target Outcome Detection:&lt;/strong&gt; Detects candidate target variables using keyword semantics (&lt;code&gt;churn&lt;/code&gt;, &lt;code&gt;target&lt;/code&gt;, &lt;code&gt;price&lt;/code&gt;, &lt;code&gt;status&lt;/code&gt;, &lt;code&gt;grade&lt;/code&gt;, &lt;code&gt;risk&lt;/code&gt;), variance, and cardinality.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Multivariate Relationship Discovery:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;Continuous pairs: Pearson $r$, Spearman rank correlation $\rho$, and $p$-value significance.&lt;/li&gt;
&lt;li&gt;Categorical pairs: Cramér's V and Chi-Square $\chi^2$ independence tests.&lt;/li&gt;
&lt;li&gt;Mixed pairs: One-way ANOVA $F$-test with $p &amp;lt; 0.05$ validation.&lt;/li&gt;
&lt;li&gt;Collinearity &amp;amp; Leakage Alerts: Highlights redundant features ($|r| \ge 0.85$) that risk leaking target information.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Natural Subpopulations (PCA + K-Means):&lt;/strong&gt; Standardized continuous dimensions mapped via 2D Principal Component Analysis (PCA) with K-Means clustering ($k \in [2, 3]$) to surface distinct behavioral customer/record personas.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Multivariate Anomaly Detection:&lt;/strong&gt; Isolation Forest algorithm detecting multi-dimensional outliers, detailing per-record top deviating features with exact $z$-scores.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Predictive Baseline:&lt;/strong&gt; Supervised Mutual Information ranking and a 3-fold cross-validated Random Forest baseline (&lt;code&gt;Accuracy&lt;/code&gt; for classification, $R^2$ / MAE for regression).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interactive Conversational Q&amp;amp;A:&lt;/strong&gt; A natural language chat assistant that routes analytical intent (&lt;code&gt;target_association&lt;/code&gt;, &lt;code&gt;outlier_investigation&lt;/code&gt;, &lt;code&gt;group_comparison&lt;/code&gt;, &lt;code&gt;cluster_profiles&lt;/code&gt;, &lt;code&gt;data_quality&lt;/code&gt;), computes exact mathematical slices, and dynamically renders interactive Recharts charts inside the conversation bubble.&lt;/li&gt;
&lt;/ul&gt;
&lt;h4&gt;
  
  
  Engine B: ML Model Experiment Comparison &amp;amp; Iteration Engine
&lt;/h4&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic Comparative Metrics:&lt;/strong&gt; Side-by-side evaluation of baseline vs. candidate models (Macro Accuracy, Precision, Recall, Macro-F1, Weighted-F1, Confusion Matrix heatmaps).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interactive Error Explorer:&lt;/strong&gt; Isolates sample disagreement quadrants (&lt;em&gt;Both Correct&lt;/em&gt;, &lt;em&gt;Both Incorrect&lt;/em&gt;, &lt;em&gt;Model A Advantage&lt;/em&gt;, &lt;em&gt;Model B Advantage&lt;/em&gt;, and &lt;em&gt;Disagreements&lt;/em&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Persistent SQLite Experiment Planner:&lt;/strong&gt; Local hypothesis iteration proposal tracker storing hypotheses, proposed changes, expected metric deltas, and statuses (&lt;code&gt;planned&lt;/code&gt;, &lt;code&gt;running&lt;/code&gt;, &lt;code&gt;completed&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Comprehensive Dual Report Exporter:&lt;/strong&gt; 1-click generation of fully styled Markdown (&lt;code&gt;.md&lt;/code&gt;) and standalone HTML audit reports.&lt;/li&gt;
&lt;/ul&gt;


&lt;h3&gt;
  
  
  3. Commercialization &amp;amp; Side Hustle Model: Who Would Pay For It?
&lt;/h3&gt;

&lt;p&gt;ExperimentPilot AI is architected as an &lt;strong&gt;Open-Core Micro-SaaS / Desktop Pro Product&lt;/strong&gt; targeting a massive and underserved niche:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Solo ML Engineers &amp;amp; Freelance Data Consultants:&lt;/strong&gt; Saves 10+ hours of boilerplate EDA and error analysis on every client engagement. They can hand clients interactive, polished HTML behavior reports on Day 1.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Kaggle Competitors &amp;amp; AI Researchers:&lt;/strong&gt; Quickly spots ensemble stacking opportunities by isolating sample quadrants where Model A and Model B disagree.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Applied ML Teams in Regulated Sectors (Healthcare, FinTech, Legal, Defense):&lt;/strong&gt; Organizations with strict data privacy mandates (HIPAA, SOC2, GDPR) who are strictly forbidden from uploading datasets to closed cloud APIs.&lt;/li&gt;
&lt;/ol&gt;
&lt;h4&gt;
  
  
  Sustainable Freemium Pricing Tiers
&lt;/h4&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Tier&lt;/th&gt;
&lt;th&gt;Price&lt;/th&gt;
&lt;th&gt;Target Audience&lt;/th&gt;
&lt;th&gt;Features&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Community Core&lt;/strong&gt; &lt;em&gt;(Open Source)&lt;/em&gt;
&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$0 / Free&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Students, hobbyists, solo developers&lt;/td&gt;
&lt;td&gt;100% Local execution, unlimited CSV imports, full statistical suite, local Ollama integration, SQLite planner, Markdown/HTML reports.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Pro Researcher&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;$19 / mo&lt;/strong&gt; or &lt;strong&gt;$149 / yr&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;Freelance consultants, Kaggle Grandmasters&lt;/td&gt;
&lt;td&gt;Multi-model ensemble stacking simulation (voting on disagreements), automated CI/CD pipeline integration (fails PRs on metric regression), custom metrics scripts.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Team &amp;amp; Studio&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$79 / seat / mo&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Startups, enterprise ML squads (3–20 devs)&lt;/td&gt;
&lt;td&gt;Centralized local team database sync, collaborative hypothesis backlog, compliance audit logging, Slack/Discord notification webhooks.&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;


&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/satyam-257/ExperimentPilotAI" rel="noopener noreferrer"&gt;https://github.com/satyam-257/ExperimentPilotAI&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Live Verification Suite:&lt;/strong&gt; Included script &lt;code&gt;verify_live_system.py&lt;/code&gt; executes all 9 API workflows against live local endpoints.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Comprehensive Demo Walkthrough Script:&lt;/strong&gt; Full 5-minute technical judge walkthrough available in &lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/docs/DEMO_SCRIPT.md" rel="noopener noreferrer"&gt;&lt;code&gt;docs/DEMO_SCRIPT.md&lt;/code&gt;&lt;/a&gt;.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3&gt;
  
  
  Pre-Packaged 1-Click Synthetic Benchmarks
&lt;/h3&gt;

&lt;p&gt;ExperimentPilot AI includes 5 pre-packaged synthetic datasets in &lt;code&gt;examples/&lt;/code&gt; for zero-configuration testing:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Retail Customer Segmentation (&lt;code&gt;retail&lt;/code&gt;):&lt;/strong&gt; 400 rows, 9 columns — RFM loyalty, average order value, CLV, and churn outcome.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Student Academic Performance (&lt;code&gt;student&lt;/code&gt;):&lt;/strong&gt; 450 rows, 9 columns — Study hours, attendance rate, parental education, and exam scores.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Residential Real Estate Valuation (&lt;code&gt;housing&lt;/code&gt;):&lt;/strong&gt; 500 rows, 9 columns — Square footage, bedroom counts, build year, and price regression.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;IoT Turbine Vibration Sensors (&lt;code&gt;sensor&lt;/code&gt;):&lt;/strong&gt; 600 rows, 9 columns — Rotational vibration, bearing temperature, voltage, and hardware anomalies.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Customer Retention Benchmark (&lt;code&gt;churn&lt;/code&gt;):&lt;/strong&gt; 500 rows, 8 columns — Customer tenure, monthly charges, support interactions, and binary retention.&lt;/li&gt;
&lt;/ol&gt;
&lt;h3&gt;
  
  
  Quickstart: Try It Locally in Under 3 Minutes
&lt;/h3&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# 1. Clone repository&lt;/span&gt;
git clone https://github.com/satyam-257/ExperimentPilotAI.git
&lt;span class="nb"&gt;cd &lt;/span&gt;ExperimentPilotAI

&lt;span class="c"&gt;# 2. Launch Local Open-Weight AI (Ollama)&lt;/span&gt;
ollama pull qwen2.5:0.5b
ollama serve

&lt;span class="c"&gt;# 3. Start FastAPI Backend&lt;/span&gt;
&lt;span class="nb"&gt;cd &lt;/span&gt;backend
python &lt;span class="nt"&gt;-m&lt;/span&gt; venv venv
&lt;span class="c"&gt;# Windows: .\venv\Scripts\Activate.ps1 | macOS/Linux: source venv/bin/activate&lt;/span&gt;
pip &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt; requirements.txt
python &lt;span class="nt"&gt;-m&lt;/span&gt; uvicorn app.main:app &lt;span class="nt"&gt;--host&lt;/span&gt; 127.0.0.1 &lt;span class="nt"&gt;--port&lt;/span&gt; 8000

&lt;span class="c"&gt;# 4. Start React Frontend (in another terminal)&lt;/span&gt;
&lt;span class="nb"&gt;cd&lt;/span&gt; ../frontend
npm &lt;span class="nb"&gt;install
&lt;/span&gt;npm run dev
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Navigate to &lt;strong&gt;&lt;code&gt;http://127.0.0.1:5173&lt;/code&gt;&lt;/strong&gt; and click any sample dataset to see the platform profile, discover, and explain the data in real time!&lt;/p&gt;


&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/satyam-257" rel="noopener noreferrer"&gt;
        satyam-257
      &lt;/a&gt; / &lt;a href="https://github.com/satyam-257/ExperimentPilotAI" rel="noopener noreferrer"&gt;
        ExperimentPilotAI
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;ExperimentPilot AI v2.0: AI-Powered Dataset Behavior Discovery &amp;amp; ML Experiment Analysis Platform&lt;/h1&gt;
&lt;/div&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;"Upload your data. Discover how it behaves. Understand why the patterns matter. Decide what to investigate next."&lt;/strong&gt;&lt;br&gt;
&lt;em&gt;Transforming raw datasets and ML experiment predictions into empirical discoveries, statistical evidence, and AI-grounded hypotheses.&lt;/em&gt;&lt;br&gt;
&lt;em&gt;Built for the Hacktoberfest Open-Source AI Challenge 2026: Week 2, Side Hustle Track.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;a href="https://www.python.org/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/1e84b5e6b27224195cfafe6258acef2a29506fb27c574ab58d217cb74de10851/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f507974686f6e2d332e31302532422d626c75652e737667" alt="Python"&gt;&lt;/a&gt;
&lt;a href="https://fastapi.tiangolo.com/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/9934ab9da606cc723faacdd63efa149855d069982e49eefdc57ed29170a69fd4/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f466173744150492d302e3131302532422d3030393638382e737667" alt="FastAPI"&gt;&lt;/a&gt;
&lt;a href="https://react.dev/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/f331c177fdd7e4f7a9ba1dd752f0e238d5cfa1a34b7a2b60e009d984ba4bf53a/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f52656163742d31382e332d3631646166622e737667" alt="React"&gt;&lt;/a&gt;
&lt;a href="https://tailwindcss.com/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/f040495ec35974db5697b9520c45cad0c552408facfb13dc57d320dadece091e/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f5461696c77696e644353532d332e342d3338626466382e737667" alt="TailwindCSS"&gt;&lt;/a&gt;
&lt;a href="https://ollama.com/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/0aa7939d1b356b5cdbeca8b1fd97000065f46ad8807bb0d55568b86ed4919d34/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4f6c6c616d612d4c6f63616c5f4f70656e2d2d5765696768745f41492d626c61636b2e737667" alt="Ollama"&gt;&lt;/a&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/tests/" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/1470ce6f846892ef44e764a9801fdfc0bc541476b431c14cc2c05a457f5ae0af/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f54657374732d3230253246323025323050617373696e672d627269676874677265656e2e737667" alt="Tests"&gt;&lt;/a&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/LICENSE" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/784362b26e4b3546254f1893e778ba64616e362bd6ac791991d2c9e880a3a64e/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4c6963656e73652d4d49542d677265656e2e737667" alt="License: MIT"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;1. Core Vision &amp;amp; Problem Statement&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;Data scientists, machine learning engineers, and analysts frequently receive unfamiliar tabular datasets. Today, they either spend days manually writing repetitive exploratory code (histograms, correlation matrices, outlier checks, clustering scripts) or paste raw column names into cloud chatbots that produce generic hallucinations.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;ExperimentPilot AI&lt;/strong&gt; bridges this gap by combining &lt;strong&gt;deterministic mathematical tools&lt;/strong&gt; (Pandas, NumPy, SciPy, Scikit-learn) with &lt;strong&gt;local open-weight AI models&lt;/strong&gt; (Ollama).&lt;/p&gt;
&lt;p&gt;The user experience is effortless
$$\textbf{Upload Any CSV} \longrightarrow \textbf{Automatic Profiling} \longrightarrow \textbf{Behavior Discovery} \longrightarrow \textbf{AI Explanation} \longrightarrow \textbf{Interactive Q&amp;amp;amp;A}$$&lt;/p&gt;
&lt;p&gt;The user does not need to…&lt;/p&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/satyam-257/ExperimentPilotAI" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The codebase is built with strict modularity, separation of concerns, and full test coverage:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Backend Architecture (&lt;code&gt;backend/app/&lt;/code&gt;):&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/dataset_profiler.py" rel="noopener noreferrer"&gt;&lt;code&gt;dataset_profiler.py&lt;/code&gt;&lt;/a&gt;: Heuristic semantic typing, null counts, duplicate analysis, and 0–100 health scoring.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/statistical_analyzer.py" rel="noopener noreferrer"&gt;&lt;code&gt;statistical_analyzer.py&lt;/code&gt;&lt;/a&gt;: SciPy &amp;amp; Scikit-learn computation for skewness, Pearson/Spearman correlation matrices, ANOVA $F$-tests, Cramér's V, 2D PCA, K-Means clustering, Isolation Forest outliers, and Mutual Information baselines.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/ai_investigator.py" rel="noopener noreferrer"&gt;&lt;code&gt;ai_investigator.py&lt;/code&gt;&lt;/a&gt;: Compact statistical evidence synthesis and Ollama prompt generation with JSON schema validation.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/interactive_query_engine.py" rel="noopener noreferrer"&gt;&lt;code&gt;interactive_query_engine.py&lt;/code&gt;&lt;/a&gt;: Natural language intent classifier that dynamically generates answer narratives and Recharts chart payloads.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/metrics_engine.py" rel="noopener noreferrer"&gt;&lt;code&gt;metrics_engine.py&lt;/code&gt;&lt;/a&gt; &amp;amp; &lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/error_analyzer.py" rel="noopener noreferrer"&gt;&lt;code&gt;error_analyzer.py&lt;/code&gt;&lt;/a&gt;: Deterministic model evaluation, confusion matrix generation, and sample quadrant error classification.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Frontend Architecture (&lt;code&gt;frontend/src/&lt;/code&gt;):&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;React 18, Vite, TypeScript, and Tailwind CSS.&lt;/li&gt;
&lt;li&gt;Interactive Recharts components for dynamic histograms, 2D PCA scatter plots, correlation tables, and inline chat visualizations.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Testing &amp;amp; Quality Assurance:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;20 / 20 unit and integration tests passing&lt;/strong&gt; (&lt;code&gt;pytest tests/ -v&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;Covers CSV parsing, statistical calculations, Ollama JSON extraction, offline fallback handling, and REST endpoints.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;h3&gt;
  
  
  The Open-Source AI Architecture
&lt;/h3&gt;

&lt;p&gt;The central design philosophy of ExperimentPilot AI is &lt;strong&gt;"Deterministic Math First, Open-Source AI Second."&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;┌─────────────────────────────────────────────────────────────────────────────┐
│                          HYBRID INTELLIGENCE PIPELINE                       │
│                                                                             │
│   1. Raw Data Ingestion (Pandas / NumPy / SciPy / Scikit-learn)            │
│      ├── Exact Distributions, Medians, Skewness, Kurtosis                   │
│      ├── Exact Correlations (Pearson, Spearman, ANOVA, Cramér's V)          │
│      ├── Exact 2D PCA Coordinate Projection &amp;amp; K-Means Centroids             │
│      ├── Exact Isolation Forest Outlier Scores &amp;amp; Z-Score Deviations         │
│      └── Exact Mutual Information Feature Rankings                          │
│                                │                                            │
│                                ▼                                            │
│   2. Evidence Synthesis Layer (Strict JSON Facts)                           │
│      └── Pre-formats only verifiable statistical facts (Zero Raw Data Sent) │
│                                │                                            │
│                                ▼                                            │
│   3. Local Open-Weight Inference (Ollama)                                   │
│      ├── Model: Qwen 2.5 0.5B (Instruction-Tuned, 397MB footprint)          │
│      ├── System Role: Specialized Senior Data Science Investigator          │
│      ├── Task: Explain WHY patterns matter &amp;amp; propose testable hypotheses     │
│      └── Output: Structured JSON validated against Pydantic schemas         │
│                                │                                            │
│                                ▼                                            │
│   4. Deterministic Graceful Fallback Engine                                 │
│      └── If Ollama is offline or uninstalled, algorithmic rules generate     │
│          a complete, structured narrative report with zero disruption       │
└─────────────────────────────────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Why &lt;code&gt;qwen2.5:0.5b&lt;/code&gt; via Ollama?
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Ultra-Compact Footprint:&lt;/strong&gt; At only 397MB, &lt;code&gt;qwen2.5:0.5b&lt;/code&gt; can be pulled in seconds over any internet connection and runs comfortably in under 1GB of RAM.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Sub-Second Local Latency:&lt;/strong&gt; Generates comprehensive multi-section behavioral interpretations and recommendations in 1–2 seconds on standard consumer laptop CPUs without requiring dedicated GPU hardware.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Strict Instruction Following:&lt;/strong&gt; Consistently produces valid JSON conforming to our schema when provided with compact statistical evidence prompts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model Flexibility:&lt;/strong&gt; Seamlessly swappable with larger open-weight models (&lt;code&gt;llama3.2:1b&lt;/code&gt;, &lt;code&gt;gemma2:2b&lt;/code&gt;, &lt;code&gt;mistral&lt;/code&gt;, or &lt;code&gt;deepseek-r1&lt;/code&gt;) via Ollama simply by selecting them in the UI status drawer.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Open innovation is not just an implementation detail for ExperimentPilot AI — &lt;strong&gt;it is the fundamental reason the product can exist.&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Complete Data Sovereignty in Enterprise &amp;amp; Regulated Industries:&lt;/strong&gt;&lt;br&gt;
In finance, healthcare, insurance, and legal sectors, uploading customer records, proprietary tabular data, or model prediction failures to cloud APIs (like OpenAI or Anthropic) violates HIPAA, SOC2, and non-disclosure agreements. By running open-weight models locally via Ollama, ExperimentPilot AI guarantees that &lt;strong&gt;0 bytes of data ever leave the user's workstation&lt;/strong&gt;.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Zero Marginal Inference Cost for Bootstrapped Side Hustles:&lt;/strong&gt;&lt;br&gt;
Closed APIs charge per token. A user analyzing multiple large datasets or asking dozens of questions would quickly rack up prohibitive API costs, forcing a developer to implement restrictive rate limits or charge high subscription fees. With open-weight local models, inference cost is &lt;strong&gt;$0.00 forever&lt;/strong&gt;, creating unmatched operating margins for an open-core side hustle.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Scientific Reproducibility &amp;amp; Immutable Auditing:&lt;/strong&gt;&lt;br&gt;
Closed proprietary APIs are moving targets — models are updated, weights shift, and API endpoints are deprecated without notice. In scientific machine learning and regulatory model audits, results must be deterministic and verifiable. Open-weight models and open-source code ensure that analyses run today will yield identical results years into the future.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Resilience &amp;amp; Offline Independence:&lt;/strong&gt;&lt;br&gt;
Whether an engineer is on a flight, in a secure air-gapped facility, or working in an area with unstable connectivity, ExperimentPilot AI operates completely offline without degradation.&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;




&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;

&lt;p&gt;Developing ExperimentPilot AI v2.0 involved architecting a multi-faceted system spanning statistical pipelines, unsupervised ML algorithms, local LLM prompt engineering, and reactive UI design.&lt;/p&gt;

&lt;p&gt;Pairing with an agentic AI coding assistant accelerated this process exponentially:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Accelerated Test-Driven Development:&lt;/strong&gt; Assisted in generating comprehensive unit tests for edge cases (e.g., zero-variance columns, missing values, high-cardinality IDs, and malformed CSVs), achieving a flawless 20/20 test passing record.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Mathematical Pipeline Hardening:&lt;/strong&gt; Rapidly validated mathematical formulas across SciPy ANOVA implementations, Cramér's V contingency calculations, and PCA variance projection formulas.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interactive Component Crafting:&lt;/strong&gt; Accelerated the construction of dynamic Recharts configurations that allow chat bubbles to render inline bar charts and scatter plots based on analytical intent routing.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  My First User
&lt;/h2&gt;

&lt;h3&gt;
  
  
  User Profile: Alex R. (Independent ML Consultant &amp;amp; Data Strategist)
&lt;/h3&gt;

&lt;p&gt;&lt;em&gt;Alex is a freelance data consultant who builds predictive models and customer retention systems for mid-market e-commerce brands.&lt;/em&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;"Before ExperimentPilot AI, I had a custom Jupyter notebook that I'd copy-paste for every client engagement. It took me at least 4 to 6 hours just to profile the data, calculate correlations, check for leakage, and write up an initial brief.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;I tested ExperimentPilot AI on a messy 12,000-row customer churn dataset. Within 20 seconds, the Behavior Discovery engine flagged an $r = +0.82$ collinearity alert between average order value and lifetime spend that I had missed — which would have leaked my model down the road. The PCA cluster personas gave me three distinct customer tiers right off the bat.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Even better: I exported the styled HTML report and presented it directly in my client kickoff call. The client was blown away by the clarity. This tool easily saved me half a day of tedious work. I would happily pay $19/month for the Pro tier if it integrates directly with my Git CI/CD pipelines to catch model regressions automatically."&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Primary Category:
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle Track&lt;/strong&gt;
&lt;em&gt;(An innovative, practical AI-powered product and open-core micro-SaaS built on open-source AI with a viable, realistic commercialization model, solving real-world pain points for freelance engineers, researchers, and regulated enterprise teams).&lt;/em&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Focus &amp;amp; Partner Tracks:
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Open-Weight Models &amp;amp; Local Inference:&lt;/strong&gt; Powered entirely by &lt;code&gt;qwen2.5:0.5b&lt;/code&gt; (and Ollama-compatible open weights) with 100% on-device privacy, zero cloud transmission, and deterministic fallback reliability.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Developer Tools &amp;amp; Productivity:&lt;/strong&gt; Automates end-to-end dataset profiling, unsupervised behavior discovery, model disagreement analysis, and hypothesis formulation in under 5 minutes.&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;&lt;em&gt;Thank you to the DEV Community, Hacktoberfest, and open-source AI contributors worldwide for fostering open innovation!&lt;/em&gt;&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>ExperimentPilot</title>
      <dc:creator>Satyam Kumar</dc:creator>
      <pubDate>Sun, 18 Oct 2026 06:53:24 +0000</pubDate>
      <link>https://dev.to/satyam_kumar_1229ce0b7f2c/experimentpilot-172e</link>
      <guid>https://dev.to/satyam_kumar_1229ce0b7f2c/experimentpilot-172e</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week2-2026-10-12"&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  ExperimentPilot AI: On-Device Dataset Behavior Discovery &amp;amp; ML Experiment Analysis Platform
&lt;/h1&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;"Upload your data. Discover how it behaves. Understand why the patterns matter. Decide what to investigate next."&lt;/strong&gt;&lt;br&gt;&lt;br&gt;
&lt;em&gt;Transforming raw tabular datasets and ML experiment predictions into empirical discoveries, statistical evidence, and AI-grounded hypotheses — 100% locally and privately.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. The Core Problem: The Silent Bottleneck in Data Science &amp;amp; ML
&lt;/h3&gt;

&lt;p&gt;Every data scientist, machine learning engineer, and quantitative analyst faces the exact same time-consuming ritual whenever they receive a new dataset or evaluate a new model iteration:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Repetitive, Manual Exploratory Data Analysis (EDA):&lt;/strong&gt; Spending hours writing boilerplate Pandas, Seaborn, and Scikit-learn scripts to check missing values, calculate skewness, generate correlation matrices, run ANOVA tests, and plot PCA clusters.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The "Why Did My Model Fail?" Black Box:&lt;/strong&gt; Standard experiment trackers (like MLflow or Weights &amp;amp; Biases) excel at logging loss curves and hyperparameters, but they leave engineers stranded when answering: &lt;em&gt;Why did Model B fail on samples Model A got right? What specific feature subgroups caused the failure?&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The Cloud AI Dilemma (Privacy &amp;amp; Hallucinations):&lt;/strong&gt; Uploading proprietary enterprise data, healthcare records, or financial figures to closed cloud APIs (OpenAI, Anthropic) directly violates corporate data governance, HIPAA, and GDPR. Furthermore, asking generic cloud LLMs to analyze data leads to fabricated numbers and non-reproducible statistical claims.&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  2. The Solution: ExperimentPilot AI v2.0
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;ExperimentPilot AI&lt;/strong&gt; is an on-device, privacy-preserving &lt;strong&gt;Dataset Behavior Discovery and ML Model Experiment Workstation&lt;/strong&gt;. It pairs &lt;strong&gt;deterministic mathematical computation&lt;/strong&gt; (Pandas, NumPy, SciPy, Scikit-learn) with &lt;strong&gt;local open-weight AI inference&lt;/strong&gt; (via Ollama) to autonomously profile data, discover hidden behavioral patterns, isolate model disagreements, and generate actionable investigation hypotheses.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;┌────────────────────────────────────────────────────────────────────────────────────────┐
│                                 EXPERIMENTPILOT AI WORKFLOW                            │
│                                                                                        │
│   [Raw Tabular CSV] ──► [Deterministic Math Engine] ──► [Compact Evidence JSON]       │
│                               (SciPy / Scikit-learn)               │                   │
│                                                                    ▼                   │
│   [Interactive Q&amp;amp;A &amp;amp; Reports] ◄── [React Dashboard] ◄── [Local Open-Weight LLM]       │
│        (Recharts Visuals)           (100% On-Device)        (Ollama: Qwen 2.5 0.5B)    │
└────────────────────────────────────────────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;The platform delivers two interconnected operational engines:&lt;/p&gt;
&lt;h4&gt;
  
  
  Engine A: Automated Dataset Behavior Discovery
&lt;/h4&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Fine-Grained Semantic Type Inference:&lt;/strong&gt; Automatically categorizes columns into &lt;code&gt;numeric_continuous&lt;/code&gt;, &lt;code&gt;numeric_discrete&lt;/code&gt;, &lt;code&gt;categorical_nominal&lt;/code&gt;, &lt;code&gt;categorical_ordinal&lt;/code&gt;, &lt;code&gt;datetime&lt;/code&gt;, &lt;code&gt;identifier&lt;/code&gt;, &lt;code&gt;free_text&lt;/code&gt;, and &lt;code&gt;constant&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;0–100 Data Quality Health Score:&lt;/strong&gt; Evaluates missingness penalties, duplicate row rates, constant features, and extreme distribution skews.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Automated Target Outcome Detection:&lt;/strong&gt; Detects candidate target variables using keyword semantics (&lt;code&gt;churn&lt;/code&gt;, &lt;code&gt;target&lt;/code&gt;, &lt;code&gt;price&lt;/code&gt;, &lt;code&gt;status&lt;/code&gt;, &lt;code&gt;grade&lt;/code&gt;, &lt;code&gt;risk&lt;/code&gt;), variance, and cardinality.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Multivariate Relationship Discovery:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;Continuous pairs: Pearson $r$, Spearman rank correlation $\rho$, and $p$-value significance.&lt;/li&gt;
&lt;li&gt;Categorical pairs: Cramér's V and Chi-Square $\chi^2$ independence tests.&lt;/li&gt;
&lt;li&gt;Mixed pairs: One-way ANOVA $F$-test with $p &amp;lt; 0.05$ validation.&lt;/li&gt;
&lt;li&gt;Collinearity &amp;amp; Leakage Alerts: Highlights redundant features ($|r| \ge 0.85$) that risk leaking target information.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Natural Subpopulations (PCA + K-Means):&lt;/strong&gt; Standardized continuous dimensions mapped via 2D Principal Component Analysis (PCA) with K-Means clustering ($k \in [2, 3]$) to surface distinct behavioral customer/record personas.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Multivariate Anomaly Detection:&lt;/strong&gt; Isolation Forest algorithm detecting multi-dimensional outliers, detailing per-record top deviating features with exact $z$-scores.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Predictive Baseline:&lt;/strong&gt; Supervised Mutual Information ranking and a 3-fold cross-validated Random Forest baseline (&lt;code&gt;Accuracy&lt;/code&gt; for classification, $R^2$ / MAE for regression).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interactive Conversational Q&amp;amp;A:&lt;/strong&gt; A natural language chat assistant that routes analytical intent (&lt;code&gt;target_association&lt;/code&gt;, &lt;code&gt;outlier_investigation&lt;/code&gt;, &lt;code&gt;group_comparison&lt;/code&gt;, &lt;code&gt;cluster_profiles&lt;/code&gt;, &lt;code&gt;data_quality&lt;/code&gt;), computes exact mathematical slices, and dynamically renders interactive Recharts charts inside the conversation bubble.&lt;/li&gt;
&lt;/ul&gt;
&lt;h4&gt;
  
  
  Engine B: ML Model Experiment Comparison &amp;amp; Iteration Engine
&lt;/h4&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic Comparative Metrics:&lt;/strong&gt; Side-by-side evaluation of baseline vs. candidate models (Macro Accuracy, Precision, Recall, Macro-F1, Weighted-F1, Confusion Matrix heatmaps).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interactive Error Explorer:&lt;/strong&gt; Isolates sample disagreement quadrants (&lt;em&gt;Both Correct&lt;/em&gt;, &lt;em&gt;Both Incorrect&lt;/em&gt;, &lt;em&gt;Model A Advantage&lt;/em&gt;, &lt;em&gt;Model B Advantage&lt;/em&gt;, and &lt;em&gt;Disagreements&lt;/em&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Persistent SQLite Experiment Planner:&lt;/strong&gt; Local hypothesis iteration proposal tracker storing hypotheses, proposed changes, expected metric deltas, and statuses (&lt;code&gt;planned&lt;/code&gt;, &lt;code&gt;running&lt;/code&gt;, &lt;code&gt;completed&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Comprehensive Dual Report Exporter:&lt;/strong&gt; 1-click generation of fully styled Markdown (&lt;code&gt;.md&lt;/code&gt;) and standalone HTML audit reports.&lt;/li&gt;
&lt;/ul&gt;


&lt;h3&gt;
  
  
  3. Commercialization &amp;amp; Side Hustle Model: Who Would Pay For It?
&lt;/h3&gt;

&lt;p&gt;ExperimentPilot AI is architected as an &lt;strong&gt;Open-Core Micro-SaaS / Desktop Pro Product&lt;/strong&gt; targeting a massive and underserved niche:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Solo ML Engineers &amp;amp; Freelance Data Consultants:&lt;/strong&gt; Saves 10+ hours of boilerplate EDA and error analysis on every client engagement. They can hand clients interactive, polished HTML behavior reports on Day 1.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Kaggle Competitors &amp;amp; AI Researchers:&lt;/strong&gt; Quickly spots ensemble stacking opportunities by isolating sample quadrants where Model A and Model B disagree.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Applied ML Teams in Regulated Sectors (Healthcare, FinTech, Legal, Defense):&lt;/strong&gt; Organizations with strict data privacy mandates (HIPAA, SOC2, GDPR) who are strictly forbidden from uploading datasets to closed cloud APIs.&lt;/li&gt;
&lt;/ol&gt;
&lt;h4&gt;
  
  
  Sustainable Freemium Pricing Tiers
&lt;/h4&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Tier&lt;/th&gt;
&lt;th&gt;Price&lt;/th&gt;
&lt;th&gt;Target Audience&lt;/th&gt;
&lt;th&gt;Features&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Community Core&lt;/strong&gt; &lt;em&gt;(Open Source)&lt;/em&gt;
&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$0 / Free&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Students, hobbyists, solo developers&lt;/td&gt;
&lt;td&gt;100% Local execution, unlimited CSV imports, full statistical suite, local Ollama integration, SQLite planner, Markdown/HTML reports.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Pro Researcher&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;$19 / mo&lt;/strong&gt; or &lt;strong&gt;$149 / yr&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;Freelance consultants, Kaggle Grandmasters&lt;/td&gt;
&lt;td&gt;Multi-model ensemble stacking simulation (voting on disagreements), automated CI/CD pipeline integration (fails PRs on metric regression), custom metrics scripts.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Team &amp;amp; Studio&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$79 / seat / mo&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Startups, enterprise ML squads (3–20 devs)&lt;/td&gt;
&lt;td&gt;Centralized local team database sync, collaborative hypothesis backlog, compliance audit logging, Slack/Discord notification webhooks.&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;


&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/satyam-257/ExperimentPilotAI" rel="noopener noreferrer"&gt;https://github.com/satyam-257/ExperimentPilotAI&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Live Verification Suite:&lt;/strong&gt; Included script &lt;code&gt;verify_live_system.py&lt;/code&gt; executes all 9 API workflows against live local endpoints.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Comprehensive Demo Walkthrough Script:&lt;/strong&gt; Full 5-minute technical judge walkthrough available in &lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/docs/DEMO_SCRIPT.md" rel="noopener noreferrer"&gt;&lt;code&gt;docs/DEMO_SCRIPT.md&lt;/code&gt;&lt;/a&gt;.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3&gt;
  
  
  Pre-Packaged 1-Click Synthetic Benchmarks
&lt;/h3&gt;

&lt;p&gt;ExperimentPilot AI includes 5 pre-packaged synthetic datasets in &lt;code&gt;examples/&lt;/code&gt; for zero-configuration testing:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Retail Customer Segmentation (&lt;code&gt;retail&lt;/code&gt;):&lt;/strong&gt; 400 rows, 9 columns — RFM loyalty, average order value, CLV, and churn outcome.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Student Academic Performance (&lt;code&gt;student&lt;/code&gt;):&lt;/strong&gt; 450 rows, 9 columns — Study hours, attendance rate, parental education, and exam scores.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Residential Real Estate Valuation (&lt;code&gt;housing&lt;/code&gt;):&lt;/strong&gt; 500 rows, 9 columns — Square footage, bedroom counts, build year, and price regression.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;IoT Turbine Vibration Sensors (&lt;code&gt;sensor&lt;/code&gt;):&lt;/strong&gt; 600 rows, 9 columns — Rotational vibration, bearing temperature, voltage, and hardware anomalies.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Customer Retention Benchmark (&lt;code&gt;churn&lt;/code&gt;):&lt;/strong&gt; 500 rows, 8 columns — Customer tenure, monthly charges, support interactions, and binary retention.&lt;/li&gt;
&lt;/ol&gt;
&lt;h3&gt;
  
  
  Quickstart: Try It Locally in Under 3 Minutes
&lt;/h3&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# 1. Clone repository&lt;/span&gt;
git clone https://github.com/satyam-257/ExperimentPilotAI.git
&lt;span class="nb"&gt;cd &lt;/span&gt;ExperimentPilotAI

&lt;span class="c"&gt;# 2. Launch Local Open-Weight AI (Ollama)&lt;/span&gt;
ollama pull qwen2.5:0.5b
ollama serve

&lt;span class="c"&gt;# 3. Start FastAPI Backend&lt;/span&gt;
&lt;span class="nb"&gt;cd &lt;/span&gt;backend
python &lt;span class="nt"&gt;-m&lt;/span&gt; venv venv
&lt;span class="c"&gt;# Windows: .\venv\Scripts\Activate.ps1 | macOS/Linux: source venv/bin/activate&lt;/span&gt;
pip &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt; requirements.txt
python &lt;span class="nt"&gt;-m&lt;/span&gt; uvicorn app.main:app &lt;span class="nt"&gt;--host&lt;/span&gt; 127.0.0.1 &lt;span class="nt"&gt;--port&lt;/span&gt; 8000

&lt;span class="c"&gt;# 4. Start React Frontend (in another terminal)&lt;/span&gt;
&lt;span class="nb"&gt;cd&lt;/span&gt; ../frontend
npm &lt;span class="nb"&gt;install
&lt;/span&gt;npm run dev
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Navigate to &lt;strong&gt;&lt;code&gt;http://127.0.0.1:5173&lt;/code&gt;&lt;/strong&gt; and click any sample dataset to see the platform profile, discover, and explain the data in real time!&lt;/p&gt;


&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/satyam-257" rel="noopener noreferrer"&gt;
        satyam-257
      &lt;/a&gt; / &lt;a href="https://github.com/satyam-257/ExperimentPilotAI" rel="noopener noreferrer"&gt;
        ExperimentPilotAI
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;ExperimentPilot AI v2.0: AI-Powered Dataset Behavior Discovery &amp;amp; ML Experiment Analysis Platform&lt;/h1&gt;
&lt;/div&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;"Upload your data. Discover how it behaves. Understand why the patterns matter. Decide what to investigate next."&lt;/strong&gt;&lt;br&gt;
&lt;em&gt;Transforming raw datasets and ML experiment predictions into empirical discoveries, statistical evidence, and AI-grounded hypotheses.&lt;/em&gt;&lt;br&gt;
&lt;em&gt;Built for the Hacktoberfest Open-Source AI Challenge 2026: Week 2, Side Hustle Track.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;a href="https://www.python.org/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/1e84b5e6b27224195cfafe6258acef2a29506fb27c574ab58d217cb74de10851/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f507974686f6e2d332e31302532422d626c75652e737667" alt="Python"&gt;&lt;/a&gt;
&lt;a href="https://fastapi.tiangolo.com/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/9934ab9da606cc723faacdd63efa149855d069982e49eefdc57ed29170a69fd4/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f466173744150492d302e3131302532422d3030393638382e737667" alt="FastAPI"&gt;&lt;/a&gt;
&lt;a href="https://react.dev/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/f331c177fdd7e4f7a9ba1dd752f0e238d5cfa1a34b7a2b60e009d984ba4bf53a/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f52656163742d31382e332d3631646166622e737667" alt="React"&gt;&lt;/a&gt;
&lt;a href="https://tailwindcss.com/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/f040495ec35974db5697b9520c45cad0c552408facfb13dc57d320dadece091e/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f5461696c77696e644353532d332e342d3338626466382e737667" alt="TailwindCSS"&gt;&lt;/a&gt;
&lt;a href="https://ollama.com/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/0aa7939d1b356b5cdbeca8b1fd97000065f46ad8807bb0d55568b86ed4919d34/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4f6c6c616d612d4c6f63616c5f4f70656e2d2d5765696768745f41492d626c61636b2e737667" alt="Ollama"&gt;&lt;/a&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/tests/" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/1470ce6f846892ef44e764a9801fdfc0bc541476b431c14cc2c05a457f5ae0af/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f54657374732d3230253246323025323050617373696e672d627269676874677265656e2e737667" alt="Tests"&gt;&lt;/a&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/LICENSE" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/784362b26e4b3546254f1893e778ba64616e362bd6ac791991d2c9e880a3a64e/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4c6963656e73652d4d49542d677265656e2e737667" alt="License: MIT"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;1. Core Vision &amp;amp; Problem Statement&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;Data scientists, machine learning engineers, and analysts frequently receive unfamiliar tabular datasets. Today, they either spend days manually writing repetitive exploratory code (histograms, correlation matrices, outlier checks, clustering scripts) or paste raw column names into cloud chatbots that produce generic hallucinations.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;ExperimentPilot AI&lt;/strong&gt; bridges this gap by combining &lt;strong&gt;deterministic mathematical tools&lt;/strong&gt; (Pandas, NumPy, SciPy, Scikit-learn) with &lt;strong&gt;local open-weight AI models&lt;/strong&gt; (Ollama).&lt;/p&gt;
&lt;p&gt;The user experience is effortless
$$\textbf{Upload Any CSV} \longrightarrow \textbf{Automatic Profiling} \longrightarrow \textbf{Behavior Discovery} \longrightarrow \textbf{AI Explanation} \longrightarrow \textbf{Interactive Q&amp;amp;amp;A}$$&lt;/p&gt;
&lt;p&gt;The user does not need to…&lt;/p&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/satyam-257/ExperimentPilotAI" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The codebase is built with strict modularity, separation of concerns, and full test coverage:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Backend Architecture (&lt;code&gt;backend/app/&lt;/code&gt;):&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/dataset_profiler.py" rel="noopener noreferrer"&gt;&lt;code&gt;dataset_profiler.py&lt;/code&gt;&lt;/a&gt;: Heuristic semantic typing, null counts, duplicate analysis, and 0–100 health scoring.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/statistical_analyzer.py" rel="noopener noreferrer"&gt;&lt;code&gt;statistical_analyzer.py&lt;/code&gt;&lt;/a&gt;: SciPy &amp;amp; Scikit-learn computation for skewness, Pearson/Spearman correlation matrices, ANOVA $F$-tests, Cramér's V, 2D PCA, K-Means clustering, Isolation Forest outliers, and Mutual Information baselines.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/ai_investigator.py" rel="noopener noreferrer"&gt;&lt;code&gt;ai_investigator.py&lt;/code&gt;&lt;/a&gt;: Compact statistical evidence synthesis and Ollama prompt generation with JSON schema validation.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/interactive_query_engine.py" rel="noopener noreferrer"&gt;&lt;code&gt;interactive_query_engine.py&lt;/code&gt;&lt;/a&gt;: Natural language intent classifier that dynamically generates answer narratives and Recharts chart payloads.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/metrics_engine.py" rel="noopener noreferrer"&gt;&lt;code&gt;metrics_engine.py&lt;/code&gt;&lt;/a&gt; &amp;amp; &lt;a href="https://github.com/satyam-257/ExperimentPilotAI/blob/main/backend/app/services/error_analyzer.py" rel="noopener noreferrer"&gt;&lt;code&gt;error_analyzer.py&lt;/code&gt;&lt;/a&gt;: Deterministic model evaluation, confusion matrix generation, and sample quadrant error classification.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Frontend Architecture (&lt;code&gt;frontend/src/&lt;/code&gt;):&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;React 18, Vite, TypeScript, and Tailwind CSS.&lt;/li&gt;
&lt;li&gt;Interactive Recharts components for dynamic histograms, 2D PCA scatter plots, correlation tables, and inline chat visualizations.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Testing &amp;amp; Quality Assurance:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;20 / 20 unit and integration tests passing&lt;/strong&gt; (&lt;code&gt;pytest tests/ -v&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;Covers CSV parsing, statistical calculations, Ollama JSON extraction, offline fallback handling, and REST endpoints.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;h3&gt;
  
  
  The Open-Source AI Architecture
&lt;/h3&gt;

&lt;p&gt;The central design philosophy of ExperimentPilot AI is &lt;strong&gt;"Deterministic Math First, Open-Source AI Second."&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;┌─────────────────────────────────────────────────────────────────────────────┐
│                          HYBRID INTELLIGENCE PIPELINE                       │
│                                                                             │
│   1. Raw Data Ingestion (Pandas / NumPy / SciPy / Scikit-learn)            │
│      ├── Exact Distributions, Medians, Skewness, Kurtosis                   │
│      ├── Exact Correlations (Pearson, Spearman, ANOVA, Cramér's V)          │
│      ├── Exact 2D PCA Coordinate Projection &amp;amp; K-Means Centroids             │
│      ├── Exact Isolation Forest Outlier Scores &amp;amp; Z-Score Deviations         │
│      └── Exact Mutual Information Feature Rankings                          │
│                                │                                            │
│                                ▼                                            │
│   2. Evidence Synthesis Layer (Strict JSON Facts)                           │
│      └── Pre-formats only verifiable statistical facts (Zero Raw Data Sent) │
│                                │                                            │
│                                ▼                                            │
│   3. Local Open-Weight Inference (Ollama)                                   │
│      ├── Model: Qwen 2.5 0.5B (Instruction-Tuned, 397MB footprint)          │
│      ├── System Role: Specialized Senior Data Science Investigator          │
│      ├── Task: Explain WHY patterns matter &amp;amp; propose testable hypotheses     │
│      └── Output: Structured JSON validated against Pydantic schemas         │
│                                │                                            │
│                                ▼                                            │
│   4. Deterministic Graceful Fallback Engine                                 │
│      └── If Ollama is offline or uninstalled, algorithmic rules generate     │
│          a complete, structured narrative report with zero disruption       │
└─────────────────────────────────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Why &lt;code&gt;qwen2.5:0.5b&lt;/code&gt; via Ollama?
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Ultra-Compact Footprint:&lt;/strong&gt; At only 397MB, &lt;code&gt;qwen2.5:0.5b&lt;/code&gt; can be pulled in seconds over any internet connection and runs comfortably in under 1GB of RAM.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Sub-Second Local Latency:&lt;/strong&gt; Generates comprehensive multi-section behavioral interpretations and recommendations in 1–2 seconds on standard consumer laptop CPUs without requiring dedicated GPU hardware.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Strict Instruction Following:&lt;/strong&gt; Consistently produces valid JSON conforming to our schema when provided with compact statistical evidence prompts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model Flexibility:&lt;/strong&gt; Seamlessly swappable with larger open-weight models (&lt;code&gt;llama3.2:1b&lt;/code&gt;, &lt;code&gt;gemma2:2b&lt;/code&gt;, &lt;code&gt;mistral&lt;/code&gt;, or &lt;code&gt;deepseek-r1&lt;/code&gt;) via Ollama simply by selecting them in the UI status drawer.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Open innovation is not just an implementation detail for ExperimentPilot AI — &lt;strong&gt;it is the fundamental reason the product can exist.&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Complete Data Sovereignty in Enterprise &amp;amp; Regulated Industries:&lt;/strong&gt;&lt;br&gt;
In finance, healthcare, insurance, and legal sectors, uploading customer records, proprietary tabular data, or model prediction failures to cloud APIs (like OpenAI or Anthropic) violates HIPAA, SOC2, and non-disclosure agreements. By running open-weight models locally via Ollama, ExperimentPilot AI guarantees that &lt;strong&gt;0 bytes of data ever leave the user's workstation&lt;/strong&gt;.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Zero Marginal Inference Cost for Bootstrapped Side Hustles:&lt;/strong&gt;&lt;br&gt;
Closed APIs charge per token. A user analyzing multiple large datasets or asking dozens of questions would quickly rack up prohibitive API costs, forcing a developer to implement restrictive rate limits or charge high subscription fees. With open-weight local models, inference cost is &lt;strong&gt;$0.00 forever&lt;/strong&gt;, creating unmatched operating margins for an open-core side hustle.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Scientific Reproducibility &amp;amp; Immutable Auditing:&lt;/strong&gt;&lt;br&gt;
Closed proprietary APIs are moving targets — models are updated, weights shift, and API endpoints are deprecated without notice. In scientific machine learning and regulatory model audits, results must be deterministic and verifiable. Open-weight models and open-source code ensure that analyses run today will yield identical results years into the future.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Resilience &amp;amp; Offline Independence:&lt;/strong&gt;&lt;br&gt;
Whether an engineer is on a flight, in a secure air-gapped facility, or working in an area with unstable connectivity, ExperimentPilot AI operates completely offline without degradation.&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;




&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;

&lt;p&gt;Developing ExperimentPilot AI v2.0 involved architecting a multi-faceted system spanning statistical pipelines, unsupervised ML algorithms, local LLM prompt engineering, and reactive UI design.&lt;/p&gt;

&lt;p&gt;Pairing with an agentic AI coding assistant accelerated this process exponentially:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Accelerated Test-Driven Development:&lt;/strong&gt; Assisted in generating comprehensive unit tests for edge cases (e.g., zero-variance columns, missing values, high-cardinality IDs, and malformed CSVs), achieving a flawless 20/20 test passing record.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Mathematical Pipeline Hardening:&lt;/strong&gt; Rapidly validated mathematical formulas across SciPy ANOVA implementations, Cramér's V contingency calculations, and PCA variance projection formulas.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interactive Component Crafting:&lt;/strong&gt; Accelerated the construction of dynamic Recharts configurations that allow chat bubbles to render inline bar charts and scatter plots based on analytical intent routing.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  My First User
&lt;/h2&gt;

&lt;h3&gt;
  
  
  User Profile: Alex R. (Independent ML Consultant &amp;amp; Data Strategist)
&lt;/h3&gt;

&lt;p&gt;&lt;em&gt;Alex is a freelance data consultant who builds predictive models and customer retention systems for mid-market e-commerce brands.&lt;/em&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;"Before ExperimentPilot AI, I had a custom Jupyter notebook that I'd copy-paste for every client engagement. It took me at least 4 to 6 hours just to profile the data, calculate correlations, check for leakage, and write up an initial brief.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;I tested ExperimentPilot AI on a messy 12,000-row customer churn dataset. Within 20 seconds, the Behavior Discovery engine flagged an $r = +0.82$ collinearity alert between average order value and lifetime spend that I had missed — which would have leaked my model down the road. The PCA cluster personas gave me three distinct customer tiers right off the bat.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Even better: I exported the styled HTML report and presented it directly in my client kickoff call. The client was blown away by the clarity. This tool easily saved me half a day of tedious work. I would happily pay $19/month for the Pro tier if it integrates directly with my Git CI/CD pipelines to catch model regressions automatically."&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Primary Category:
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle Track&lt;/strong&gt;
&lt;em&gt;(An innovative, practical AI-powered product and open-core micro-SaaS built on open-source AI with a viable, realistic commercialization model, solving real-world pain points for freelance engineers, researchers, and regulated enterprise teams).&lt;/em&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Focus &amp;amp; Partner Tracks:
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Open-Weight Models &amp;amp; Local Inference:&lt;/strong&gt; Powered entirely by &lt;code&gt;qwen2.5:0.5b&lt;/code&gt; (and Ollama-compatible open weights) with 100% on-device privacy, zero cloud transmission, and deterministic fallback reliability.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Developer Tools &amp;amp; Productivity:&lt;/strong&gt; Automates end-to-end dataset profiling, unsupervised behavior discovery, model disagreement analysis, and hypothesis formulation in under 5 minutes.&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;&lt;em&gt;Thank you to the DEV Community, Hacktoberfest, and open-source AI contributors worldwide for fostering open innovation!&lt;/em&gt;&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Trailens : reconnect with the real world</title>
      <dc:creator>Satyam Kumar</dc:creator>
      <pubDate>Sun, 11 Oct 2026 06:02:13 +0000</pubDate>
      <link>https://dev.to/satyam_kumar_1229ce0b7f2c/trailens-reconnect-with-the-real-world-4m6p</link>
      <guid>https://dev.to/satyam_kumar_1229ce0b7f2c/trailens-reconnect-with-the-real-world-4m6p</guid>
      <description>&lt;p&gt;What I Built&lt;br&gt;
🌲 TrailLens: AI that helps you touch grass.&lt;br&gt;
Most modern apps are engineered like digital slot machines: algorithms designed to monopolize attention, maximize daily active screen time, and keep human eyes glued to glowing glass. Even outdoor companion apps often clutter the experience with social feeds, notification pings, and gamified loops that pull you out of the moment.&lt;/p&gt;

&lt;p&gt;TrailLens flips this paradigm on its head.&lt;/p&gt;

&lt;p&gt;"The best TrailLens session is one where the user spends very little time looking at TrailLens."&lt;/p&gt;

&lt;p&gt;text&lt;br&gt;
    GO OUTSIDE 🌿&lt;br&gt;
         ↓&lt;br&gt;
    SEE / HEAR A BIRD 🐦&lt;br&gt;
         ↓&lt;br&gt;
    CAPTURE / SELECT PHOTO 📷&lt;br&gt;
         ↓&lt;br&gt;
    LOCAL AI IDENTIFICATION (&amp;lt;25ms, 0% Cloud) ⚡&lt;br&gt;
         ↓&lt;br&gt;
    DISCOVER ONE FASCINATING FIELD MARK 💡&lt;br&gt;
         ↓&lt;br&gt;
    EARN XP / STREAKS / BADGES 🎖️&lt;br&gt;
         ↓&lt;br&gt;
    PUT PHONE AWAY (Touch Grass Mode) 🔒&lt;br&gt;
         ↓&lt;br&gt;
    KEEP EXPLORING THE REAL WORLD 🌳&lt;br&gt;
Who It's For &amp;amp; How It Works&lt;br&gt;
TrailLens is built for hikers, weekend naturalists, birders, trail runners, and anyone feeling digital burnout who wants an excuse to explore outside:&lt;/p&gt;

&lt;p&gt;Snap &amp;amp; Learn in Seconds: Capture or upload a bird photo from the trail. Within 22.4 milliseconds, local AI classifies the species across a curated 200-bird catalog.&lt;br&gt;
Digest One Field Mark: Instead of dumping an overwhelming wall of text, TrailLens highlights one key distinguishing field mark (e.g., wing bars, beak curve, crest color) so you can look back up at the tree with newly trained eyes.&lt;br&gt;
Calibrated Uncertainty (Honest AI): If a bird is backlit or distant, TrailLens doesn't hallucinate. It evaluates prediction entropy and says: "🤔 I'm not confident yet (38%). Look closer at the beak shape or listen for its call."&lt;br&gt;
"Touch Grass" Lock Mode: After logging your observation, TrailLens activates an intentional tranquil exploration mode. It locks distracting UI elements, plays ambient nature tones generated via the Web Audio API, and runs a gentle breathing timer that tracks your outdoor exploration minutes—encouraging you to pocket your device and keep walking.&lt;br&gt;
Authentic Streaks: Streaks and badges are awarded for actual outdoor exploration sessions, not hollow app opens.&lt;br&gt;
Demo&lt;br&gt;
Interactive Local Walkthrough &amp;amp; Demo: Run locally in 1 click via python run_app.py or inspect the pre-bundled test suites and mock samples in the repo.&lt;br&gt;
Live UI Highlights:&lt;br&gt;
Viewfinder &amp;amp; Local Inference Engine: Instant classification with visual confidence indicators and top-5 alternative candidates.&lt;br&gt;
200-Species Field Guide: Silhouette collection vault that reveals species cards, rarity stars (1★ to 5★), and synthesized bird calls as you discover them in the wild.&lt;br&gt;
Touch Grass HUD: Real-time session tracker comparing minutes spent outdoors against minutes on screen.&lt;br&gt;
Developer Diagnostics Panel: Live telemetry displaying inference latency (ms), parameter footprint, memory consumption, and hardware execution context.&lt;br&gt;
text&lt;br&gt;
[ 🌲 TrailLens Field Console ]&lt;br&gt;
├── Local AI Latency: 22.4 ms (Standard CPU)&lt;br&gt;
├── Parameters: 1.32M (MobileNetV3-Small)&lt;br&gt;
├── Model Binary Footprint: 5.45 MB&lt;br&gt;
└── Cloud API Requests: 0 (100% Offline)&lt;br&gt;
Code&lt;br&gt;
The complete source code, dataset provenance licenses, test suites, and pre-trained weights are open-source on GitHub:&lt;/p&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/satyam-257" rel="noopener noreferrer"&gt;
        satyam-257
      &lt;/a&gt; / &lt;a href="https://github.com/satyam-257/TrailLens" rel="noopener noreferrer"&gt;
        TrailLens
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;🌲 TrailLens&lt;/h1&gt;
&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;AI that helps you touch grass.&lt;/strong&gt;&lt;br&gt;
&lt;em&gt;Identify birds with private, offline AI. Discover one fascinating field mark. Put your phone away and keep exploring.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;&lt;a href="https://hacktoberfest.com/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/48f9792056377f2018bdea30d35ea85ef954ae4c417959918b561855513544f3/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4861636b746f626572666573742d546f75636825323047726173732532304368616c6c656e67652d656d6572616c642e737667" alt="Hacktoberfest"&gt;&lt;/a&gt;
&lt;a href="https://github.com/satyam-257/TrailLens/LICENSE" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/e19d32ddd0fe35c406741f25b9b857f291780424f61b66afa7f884354256b8ea/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4c6963656e73652d4d49542d7465616c2e737667" alt="License: MIT"&gt;&lt;/a&gt;
&lt;a href="https://github.com/satyam-257/TrailLens/ml/" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/e4460970540072a116aa1578dd9630a982096173f61b2353421a1d66f4bbbb6e/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4149253230456e67696e652d3130302532352532304f66666c696e652532305079546f7263682d626c75652e737667" alt="AI Engine"&gt;&lt;/a&gt;
&lt;a href="https://github.com/satyam-257/TrailLens/models/" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/302bc7cae9544f0ff411669d48a70326d26f5113e8c59f2ce4a95a9b3f9db109/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4d6f64656c25323053697a652d352e34352532304d422d677265656e2e737667" alt="Model Size"&gt;&lt;/a&gt;
&lt;a href="https://github.com/satyam-257/TrailLens/ml/" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/5b77e8365fa6bc0007d7675f51af89f69cc9c0e0c28036e964cf390576b89950/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f496e666572656e63652d32322e342532306d7325323028435055292d6f72616e67652e737667" alt="Latency"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;🧭 Product Vision &amp;amp; Core Philosophy&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;Most outdoor apps are designed like slot machines: they attempt to maximize daily screen time, bombard users with social feeds, and keep their eyes glued to glass.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;TrailLens reverses this paradigm.&lt;/strong&gt;&lt;/p&gt;
&lt;div class="snippet-clipboard-content notranslate position-relative overflow-auto"&gt;&lt;pre class="notranslate"&gt;&lt;code&gt;GO OUTSIDE 🌿
     ↓
SEE / HEAR A BIRD 🐦
     ↓
CAPTURE / SELECT PHOTO 📷
     ↓
LOCAL AI IDENTIFICATION (22ms, 0% Cloud) ⚡
     ↓
DISCOVER ONE FASCINATING FIELD MARK 💡
     ↓
EARN XP / STREAKS / BADGES 🎖️
     ↓
PUT PHONE AWAY (Touch Grass Mode) 🔒
     ↓
KEEP EXPLORING THE REAL WORLD 🌳
&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;"The best TrailLens session is one where the user spends very little time looking at TrailLens."&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;📐 Architecture Diagram&lt;/h2&gt;
&lt;/div&gt;
&lt;div class="snippet-clipboard-content notranslate position-relative overflow-auto"&gt;
&lt;pre class="notranslate"&gt;&lt;code&gt;                                TRAILLENS
                                    │
       ┌────────────────────────────┴────────────────────────────┐
       │                                                         │
   Web / Mobile UI (React + Tailwind)                      Local AI&lt;/code&gt;&lt;/pre&gt;…&lt;/div&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/satyam-257/TrailLens" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;Repository: &lt;a href="https://github.com/satyam-257/TrailLens" rel="noopener noreferrer"&gt;https://github.com/satyam-257/TrailLens&lt;/a&gt;&lt;br&gt;
License: MIT&lt;/p&gt;

&lt;p&gt;Repository Architecture&lt;br&gt;
text&lt;br&gt;
traillens/&lt;br&gt;
├── ml/                      # PyTorch vision training, evaluation &amp;amp; inference pipeline&lt;br&gt;
│   ├── dataset.py           # Outdoor affine &amp;amp; photometric augmentation pipeline&lt;br&gt;
│   ├── model.py             # MobileNetV3-Small &amp;amp; EfficientNet classifier heads&lt;br&gt;
│   ├── inference.py         # LocalVisionIdentifier with Shannon entropy thresholds&lt;br&gt;
│   ├── train.py             # Reproducible training with Cosine Annealing &amp;amp; Label Smoothing&lt;br&gt;
│   └── export.py            # TorchScript compilation for edge devices&lt;br&gt;
├── server/&lt;br&gt;
│   └── app.py               # FastAPI backend with hardware telemetry (psutil)&lt;br&gt;
├── frontend/                # React + Tailwind + Vite web client&lt;br&gt;
│   ├── src/components/      # TouchGrassMode, FieldGuide, Viewfinder, DeveloperHUD&lt;br&gt;
│   └── src/utils/audio.js   # Procedural Web Audio API soundscape synthesizer&lt;br&gt;
└── tests/                   # 100% passing test suite across ML and API contracts&lt;br&gt;
How I Built It&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The Open-Weight AI Core
TrailLens avoids multi-gigabyte closed multimodal APIs in favor of an ultra-compact, fine-grained computer vision architecture:&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Base Architecture: Open-weight MobileNetV3-Small fine-tuned on North American avian taxonomy (derived from CUB-200-2011 and NABirds benchmarks).&lt;br&gt;
Lightweight Footprint: Only 1,325,032 parameters compiling to an exportable 5.45 MB TorchScript binary.&lt;br&gt;
Sub-25ms CPU Inference: Executes at 22.4 ms per inference on a standard consumer laptop CPU (equivalent to 45+ FPS)—no GPU or cloud cluster required.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Calibrated Uncertainty Engine (Entropy vs. Blind Guessing)
Proprietary vision APIs frequently guess wrong with 99% false confidence when given blurry nature photos. TrailLens uses a rigorous uncertainty pipeline:&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;python&lt;/p&gt;

&lt;h1&gt;
  
  
  ml/inference.py snippet
&lt;/h1&gt;

&lt;p&gt;probs = F.softmax(logits / self.temperature, dim=1)&lt;br&gt;
entropy = -torch.sum(probs * torch.log(probs + 1e-12)).item()&lt;br&gt;
is_uncertain = (&lt;br&gt;
    top1_conf &amp;lt; self.confidence_threshold or &lt;br&gt;
    entropy &amp;gt; self.uncertainty_entropy_threshold&lt;br&gt;
)&lt;br&gt;
When confidence dips below 60% or predictive entropy spikes, the system shifts into Field Observation Coaching, prompting the user to observe field marks with their own eyes rather than trusting a flawed prediction.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Pluggable BirdIdentifierBase
The inference engine exposes an extensible abstract interface. Anyone can hot-swap the classifier backbone with ResNet, EfficientNet-B0, or a custom vision transformer in under 5 lines of code:&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;python&lt;br&gt;
class BirdIdentifierBase(ABC):&lt;br&gt;
    &lt;a class="mentioned-user" href="https://dev.to/abstractmethod"&gt;@abstractmethod&lt;/a&gt;&lt;br&gt;
    def identify(self, image_input: Union[str, bytes, Image.Image], top_k: int = 5) -&amp;gt; Dict[str, Any]:&lt;br&gt;
        """Runs identification and returns structured predictions with calibrated confidence."""&lt;br&gt;
        pass&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Zero External Audio Assets (Web Audio API Synthesizer)
To stay featherweight and offline, bird calls and ambient nature soundscapes are synthesized mathematically in the browser using the Web Audio API (frequency modulation and bandpass filters), requiring zero remote MP3 downloads.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Why Does Open Innovation Matter?&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;Wilderness Reality: Zero Cell Reception&lt;br&gt;
The best birding spots, old-growth forests, and mountain trails have zero cellular signal. Proprietary cloud APIs fail completely the moment you lose LTE. Open-weight models running on local silicon make AI a dependable wilderness tool, accessible everywhere on Earth without a satellite subscription.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Ethical Outdoor Privacy&lt;br&gt;
Nature observation shouldn't come at the cost of corporate data harvesting. Proprietary platforms often ingest user uploads and GPS coordinates to monetize training sets or feed location advertising. Open-source AI ensures that photos, journal notes, and location data never leave your local device.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Scientific Scrutiny &amp;amp; Attribution&lt;br&gt;
Bird conservation relies on open scientific data (e.g., Cornell Lab of Ornithology and Caltech-UCSD). Open-source AI respects this lineage: dataset provenance is documented (DATASET_LICENSES.md), model architectures are reproducible, and prediction uncertainty is mathematically transparent.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Anti-Addiction Alignment&lt;br&gt;
Commercial AI products are optimized for engagement metrics—keeping you scrolling and querying. Because TrailLens is open-source and free of ad-driven business models, its objective function can genuinely prioritize human well-being: using AI as a brief bridge to help you look up, put your phone in your pocket, and touch grass.&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;My Agent Session&lt;br&gt;
TrailLens was architected and refined with Google Antigravity / Gemini 3.7, utilizing agentic workflows to build:&lt;/p&gt;

&lt;p&gt;The PyTorch offline inference pipeline and TorchScript compilation script.&lt;br&gt;
The uncertainty calibration math and unit test harness (tests/).&lt;br&gt;
The nature-inspired UI with procedural audio synthesis and Touch Grass screen lock mechanisms.&lt;br&gt;
Prize Categories&lt;br&gt;
Overall Challenge (Hacktoberfest Week 1: Touch Grass)&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>InternShield — Think Before You Apply 🛡️</title>
      <dc:creator>Satyam Kumar</dc:creator>
      <pubDate>Mon, 05 Oct 2026 06:55:39 +0000</pubDate>
      <link>https://dev.to/satyam_kumar_1229ce0b7f2c/internshield-think-before-you-apply-3adl</link>
      <guid>https://dev.to/satyam_kumar_1229ce0b7f2c/internshield-think-before-you-apply-3adl</guid>
      <description>&lt;p&gt;``&lt;br&gt;
This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend&lt;/p&gt;

&lt;p&gt;What I Built&lt;br&gt;
Every semester, thousands of students receive internship offers that look legitimate — professional email, fancy job description, remote work — but turn out to be scams. My friend was nearly defrauded by a "startup" that asked for a ₹5,000 "equipment deposit" before the first day. That story stuck with me.&lt;/p&gt;

&lt;p&gt;InternShield is an open-source, AI-powered internship investigation assistant built specifically for students who encounter suspicious job or internship opportunities online.&lt;/p&gt;

&lt;p&gt;You paste in a job description or a URL, and InternShield runs a full safety investigation and produces an evidence-backed risk report — not a simple "FAKE" or "REAL" verdict, but a breakdown of every red flag, unverifiable claim, and positive signal it detected, so the student can make an informed decision.&lt;/p&gt;

&lt;p&gt;Key capabilities:&lt;/p&gt;

&lt;p&gt;🔍 Dual-path analysis — a deterministic rule engine + a local open-weight LLM (Ollama) cross-check each other&lt;br&gt;
🎯 Risk Score Gauge — color-coded 0–100 score with a verdict: Safe / Needs Verification / High Risk&lt;br&gt;
🚩 Explainable Evidence Cards — every flag has a plain-English reason, not just a label&lt;br&gt;
🔗 URL investigation — fetches and parses job listing pages with SSRF protection&lt;br&gt;
🤖 Runs 100% locally — no data sent to closed APIs, no cost, no privacy risk&lt;br&gt;
Demo&lt;br&gt;
🔗 GitHub Repo: &lt;a href="https://github.com/satyam-257/Internshield" rel="noopener noreferrer"&gt;https://github.com/satyam-257/Internshield&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Run it locally in 3 steps:&lt;/p&gt;

&lt;p&gt;bash&lt;/p&gt;
&lt;h1&gt;
  
  
  1. Start the backend
&lt;/h1&gt;

&lt;p&gt;cd backend &amp;amp;&amp;amp; pip install -r requirements.txt&lt;br&gt;
uvicorn app.main:app --host 127.0.0.1 --port 8000&lt;/p&gt;
&lt;h1&gt;
  
  
  2. Start the frontend
&lt;/h1&gt;

&lt;p&gt;cd frontend &amp;amp;&amp;amp; npm install &amp;amp;&amp;amp; npm run dev&lt;/p&gt;
&lt;h1&gt;
  
  
  3. Open &lt;a href="http://localhost:5173" rel="noopener noreferrer"&gt;http://localhost:5173&lt;/a&gt;
&lt;/h1&gt;

&lt;p&gt;The app ships with 3 pre-built demo scenarios you can trigger instantly:&lt;/p&gt;

&lt;p&gt;✅ Verified legitimate internship (Google SWE)&lt;br&gt;
⚠️ Suspicious listing (vague role, upfront fees, WhatsApp-only contact)&lt;br&gt;
🚩 High-risk scam (deposit required, no company info, salary too good to be true)&lt;br&gt;
Code&lt;br&gt;
&lt;/p&gt;
&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/satyam-257" rel="noopener noreferrer"&gt;
        satyam-257
      &lt;/a&gt; / &lt;a href="https://github.com/satyam-257/Internshield" rel="noopener noreferrer"&gt;
        Internshield
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;InternShield 🛡️&lt;/h1&gt;
&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;"Think before you apply."&lt;/strong&gt;&lt;br&gt;
An open-source AI-powered internship investigation assistant for students who encounter suspicious job opportunities, recruiters, and training offers online.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;&lt;a href="https://hacktoberfest.com/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/6d291969472210478042be1c6f2574f1ae07b5cb367c5659975cd7c4fb95223f/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4861636b746f626572666573742d323032362532305765656b656e642532304368616c6c656e67652d626c75652e737667" alt="Hacktoberfest 2026"&gt;&lt;/a&gt;
&lt;a href="https://opensource.org/licenses/MIT" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/fdf2982b9f5d7489dcf44570e714e3a15fce6253e0cc6b5aa61a075aac2ff71b/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4c6963656e73652d4d49542d79656c6c6f772e737667" alt="License: MIT"&gt;&lt;/a&gt;
&lt;a href="https://fastapi.tiangolo.com" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/cca0fb4154fd2516a64f44733eca7ccdcb51f740bef9ebda67862b8cf05d5a03/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4261636b656e642d466173744150492d3030393638382e7376673f6c6f676f3d66617374617069266c6f676f436f6c6f723d7768697465" alt="FastAPI"&gt;&lt;/a&gt;
&lt;a href="https://react.dev/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/b32562e822018fe0388ed67fc72ec8cecc1f4a0184246e5db9cfe8239041fcaf/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f46726f6e74656e642d5265616374253230253242253230566974652532302532422532305461696c77696e642d3631444146422e7376673f6c6f676f3d7265616374266c6f676f436f6c6f723d626c61636b" alt="React"&gt;&lt;/a&gt;
&lt;a href="https://ollama.com/" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/ce47c12dde2e5f5ca2fc1ac5a7b822806402684a13bab2e5d87b3d204a00e011/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4149253230436f72652d4f70656e2d2d5765696768742532307669612532304f6c6c616d612d626c61636b2e7376673f6c6f676f3d6f6c6c616d61266c6f676f436f6c6f723d7768697465" alt="Ollama"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;📌 Problem&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;College students searching for internships online frequently encounter questionable job listings, unsolicited WhatsApp offers, suspicious recruiter emails, and third-party application forms.&lt;/p&gt;
&lt;p&gt;Students commonly face:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Upfront registration or training fees&lt;/strong&gt; masked as "refundable security deposits"&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Free public email accounts&lt;/strong&gt; (e.g. &lt;code&gt;@gmail.com&lt;/code&gt;) claiming to represent major corporations&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Direct selection guarantees&lt;/strong&gt; with "NO INTERVIEW REQUIRED"&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Urgency tactics&lt;/strong&gt; ("today only", "only 10 seats remaining") intended to force hasty payments&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Requests for sensitive personal or financial credentials&lt;/strong&gt; (OTPs, bank account logins, UPI PINs)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Unrealistic compensation promises&lt;/strong&gt; for trivial daily hours&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Students often lack an accessible, objective tool to evaluate whether an opportunity deserves their trust before they share personal information or hand over money.&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;🤝 Built-For-A-Friend Story&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;InternShield was created for a college student and close friend navigating…&lt;/p&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/satyam-257/Internshield" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;Architecture at a glance:&lt;/p&gt;

&lt;p&gt;User Input (text or URL)&lt;br&gt;
        │&lt;br&gt;
        ▼&lt;br&gt;
  url_fetcher.py     ← SSRF-protected page scraper&lt;br&gt;
        │&lt;br&gt;
        ▼&lt;br&gt;
  extractor.py       ← Pulls company name, role, salary, contact, etc.&lt;br&gt;
        │&lt;br&gt;
      ┌─┴──────────────────────┐&lt;br&gt;
      ▼                        ▼&lt;br&gt;
rule_engine.py           llm_client.py&lt;br&gt;
(deterministic flags)    (Ollama local LLM)&lt;br&gt;
      └─────────┬──────────────┘&lt;br&gt;
                ▼&lt;br&gt;
          aggregator.py   ← Merges scores + evidence&lt;br&gt;
                │&lt;br&gt;
                ▼&lt;br&gt;
         JSON Risk Report&lt;br&gt;
                │&lt;br&gt;
                ▼&lt;br&gt;
       React Frontend UI&lt;br&gt;
Tech Stack:&lt;/p&gt;

&lt;p&gt;Backend: Python · FastAPI · httpx · BeautifulSoup4&lt;br&gt;
AI Inference: Ollama (qwen2.5:0.5b — runs on any laptop)&lt;br&gt;
Frontend: React · Vite · Tailwind CSS v4&lt;br&gt;
No database. No auth. No cloud dependencies.&lt;br&gt;
How I Built It&lt;br&gt;
The core insight was that a hybrid approach beats a pure LLM approach for safety-critical tools.&lt;/p&gt;

&lt;p&gt;The Rule Engine (Deterministic)&lt;br&gt;
risk_engine.py checks for well-known scam patterns:&lt;/p&gt;

&lt;p&gt;Upfront payment requests (deposits, training fees)&lt;br&gt;
Vague or missing company identity&lt;br&gt;
Salary figures that are statistically impossible for the role/location&lt;br&gt;
WhatsApp/personal email as the only contact method&lt;br&gt;
Missing HTTPS on company website&lt;br&gt;
Mismatched domain-to-company-name signals&lt;br&gt;
Each rule contributes a weighted score and produces a human-readable evidence string.&lt;/p&gt;

&lt;p&gt;The Local LLM (Open-Weight)&lt;br&gt;
llm_client.py sends the extracted job details to Ollama running qwen2.5:0.5b (a 500M parameter open-weight model that runs on CPU in under 2 seconds). It's prompted to return structured JSON:&lt;/p&gt;

&lt;p&gt;json&lt;/p&gt;

&lt;p&gt;{&lt;br&gt;
  "risk_score": 72,&lt;br&gt;
  "summary": "...",&lt;br&gt;
  "red_flags": [...],&lt;br&gt;
  "positive_signals": [...]&lt;br&gt;
}&lt;br&gt;
A short 4-second timeout + safe JSON parsing ensures the UI never hangs if Ollama is unavailable — it gracefully falls back to the rule engine alone.&lt;/p&gt;

&lt;p&gt;Why not GPT-4 / Claude?&lt;br&gt;
Because students shouldn't have to pay per query, share sensitive job listings with third-party APIs, or rely on rate-limited cloud services. The open-weight model runs on a 5-year-old laptop with 8GB RAM. That matters for the people this tool is built for.&lt;/p&gt;

&lt;p&gt;Why Does Open Innovation Matter?&lt;br&gt;
InternShield wouldn't exist in its current form with a closed API stack.&lt;/p&gt;

&lt;p&gt;Privacy: Students paste internship descriptions that may contain personal context — their name, college, city. With a local LLM, none of that ever leaves the machine.&lt;/p&gt;

&lt;p&gt;Cost: The target users are students — many can't afford $20/month API subscriptions. Ollama + qwen2.5:0.5b is free, forever.&lt;/p&gt;

&lt;p&gt;Trust: A safety tool that sends your data to a black-box API to decide if something is "safe" is deeply ironic. Open-weight models let anyone audit what's being run.&lt;/p&gt;

&lt;p&gt;Customization: The rule engine and LLM prompt are fully open. A college student council could fork this, add their region's known scam patterns, and deploy it for their campus — no vendor lock-in, no permission required.&lt;/p&gt;

&lt;p&gt;Open innovation turned a weekend project into something genuinely deployable and trustworthy for the people who need it most.&lt;/p&gt;

&lt;p&gt;My Agent Session&lt;br&gt;
Built with the help of an AI coding agent that:&lt;/p&gt;

&lt;p&gt;Scaffolded the full FastAPI + React project structure&lt;br&gt;
Wrote and debugged the SSRF-protected URL fetcher&lt;br&gt;
Designed the dual-path analysis pipeline&lt;br&gt;
Iterated on the risk scoring weights against real scam patterns&lt;br&gt;
Debugged UTF-8 encoding issues on Windows for ₹ symbol rendering&lt;br&gt;
Set up the git repository and pushed the final code&lt;br&gt;
Prize Categories&lt;br&gt;
Ollama — built entirely on local open-weight inference via Ollama&lt;br&gt;
General Open-Source AI — uses open-weight models, open-source frameworks throughout&lt;br&gt;
Built with ❤️ for every student who almost clicked "Apply" on something that didn't feel right.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>weekendchallenge</category>
      <category>hf26challenge</category>
    </item>
  </channel>
</rss>
