DEV Community: Dhiraj Shirse

Sleep Peacefully. Wake Up Informed: How I Built Sloth Digest with Local AI

Dhiraj Shirse — Mon, 26 Oct 2026 07:26:01 +0000

*This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

Modern communication tools have turned our mornings into an anxiety trap. Between student councils, hackathon teams, and family WhatsApp groups, waking up often means confronting 300 to 500 unread messages: inside jokes, reaction emoji, pizza memes, and fragmented logistics. Finding whether someone actually asked you a question or set a deadline requires 35 to 45 minutes of compulsive, thumb-numbing scrolling.

I built Sloth Digest — an offline-first, local-first web application that reads your busy group chats overnight on your own computer so you wake up completely calm.

Instead of a chaotic wall of noise, you get a serene morning hearth answering four core questions:

Executive Morning Digest: A calm two-line summary plus an expandable Detailed Discussion Breakdown that synthesizes the narrative, key debates, agreements, and next steps.
AI Chat & Message Classifier: Every message is categorized as ⭐ Important (action items, deadlines, questions, blockers) or 🛡️ Spam / Banter (emojis, reactions, media placeholders, greetings), alongside an overall chat-level priority badge.
Formatted Chat Viewer: An interactive chat-bubble stream with sender avatar initials, timestamp tags, line citations (#L42), and real-time filter tabs (All, Important Only, Spam / Chatter).
Who is Waiting on You: Only the action items and questions directed at you or your team.
Dates & Deadlines Mentioned: A chronological timeline of commitments.
The Safe to Ignore Counter: A celebratory badge highlighting how much chatter you avoided (e.g., "384 messages safely bypassed").

Every extracted fact includes our Hallucination Guard: a 1-click citation linking back to the exact line number, author, timestamp, and raw text of the original chat export. There are zero cloud API calls — nothing leaves your machine.

Demo

GitHub Repository: https://github.com/Dhiraj201226/Sloth-Digest
URL: https://sloth-digest.vercel.app/
Docker 1-Command Startup: docker compose up --build -d (Local container on http://localhost:8080)
Vercel Deployable: Pre-configured with vercel.json for 1-click UI deployment.

The 30-Second Lazy Workflow:

Drop your exported chat (.txt or .zip) into the dropzone or watched_chats/ folder.
Watch the streaming progress bar report chunk-by-chunk with honest remaining time estimates.
Wake up to instant visual calm: warm linen theme (#FDFBF7), zero red alarm badges, and a full breakdown of what needs your attention.
Toggle between the ☕ Morning Digest and the 💬 Formatted Chat Stream to inspect message classifications in place.

Code

Dhiraj201226 / Sloth-Digest

🦥 Sloth Digest

Slow and steady: rewards laziness by reading busy group chats overnight on your local machine so you wake up completely calm.
Built for the Dev.to Hacktoberfest Open-Source AI Challenge (Week 3 Theme: "Slow and Steady").

The Problem

You belong to 15 busy group chats: hackathon teams, college project channels, student councils, and family groups. Over an evening, hundreds of messages accumulate: inside jokes, memes, fragmented planning, and reaction emoji.

You face a stressful choice:

The Compulsive Catch-Up: Spend 40 minutes scrolling through trivial banter just to find if someone asked you a question.
The Passive Neglect: Mute everything and risk missing a crucial deadline or assignment.

The Solution: Radical Laziness

Do nothing. Go to sleep.

Sloth Digest processes your exported WhatsApp chats (.txt or .zip) overnight on your local machine using an open-weight model via Ollama. When you wake up, a serene…

View on GitHub

Tech Stack Highlights:

Local AI Core: Ollama running qwen2.5:1.5b (configurable via .env or UI to llama3.2:3b or gemma4:12b).
Backend: Python 3.10+ with FastAPI, Pydantic v2 schemas, and embedded SQLite for zero-config persistence.
Frontend: Vite + React 19 + TypeScript + Tailwind CSS v4, built to strict WCAG 2.2 AA accessibility standards.
Automated Airgap Test: Automated Pytest asserting that any socket connection attempting to leave 127.0.0.1 immediately fails the build.

How I Built It

I approached Sloth Digest as a pairing of HCI design discipline and systems engineering across 7 deliberate phases:

HCI Personas & Usability Specifications (docs/01_requirements.md): Designed around three distinct cognitive, perception, and motor profiles (an overwhelmed CS student, a busy class representative, and a senior family coordinator). Established 10 quantitative usability specifications (U1–U10) before writing code.
The "Calm Morning Hearth" Design (docs/02_design.md): Replaced conventional urgent unread counters and alarm badges with warm linen tones (#FDFBF7), sage accents (#4A6B53), and generous spacing.
Map-Reduce Orchestration (docs/03_architecture.md): Because chat exports often exceed 5,000+ messages, a two-stage map-reduce engine chunks chats into temporal slices (50–60 messages with overlaps), extracts candidate facts in parallel, and reduces them into a calm executive summary.
Hallucination Guard & Citation Anchoring: Built an $O(1)$ in-memory index that anchors extracted action items and deadlines back to raw line numbers. Unanchored model output is discarded.
Nightly Job (sloth run) & File Tidier: Added a background CLI runner with SHA256 file hashing for lazy overnight execution, plus an Inbox Tidier stretch goal that standardizes messy WhatsApp export filenames in 1 click.
Chat & Message Classification Engine: Implemented message-level semantic analysis categorizing chatter, reactions, and media notices from direct questions and actionable commitments.
Production Containerization: Multi-stage Dockerfile and docker-compose.yml bundling both the compiled React frontend and Python backend for 1-command deployment anywhere.

Why Does Open Innovation Matter?

Open-source and local-first AI are not just engineering preferences; they represent a fundamental ethical standard for personal software:

Private Data Stays on Your Machine: Group chats contain deeply private conversations: personal finances, family health updates, academic disputes, and passwords. Asking users to upload their personal family logs to third-party cloud APIs (OpenAI or Anthropic) is a non-starter.
Zero Running Cost: Anyone with a standard 8 GB laptop can run Sloth Digest forever without subscription fees, credit cards, or token meters.
Swappable Open Weights: Users aren't locked into a single proprietary provider. Today it's qwen2.5:1.5b; tomorrow you can swap to llama3.2:3b or gemma4:12b with a single config flag.
Reliability in the Wild: Works in remote off-grid locations, on airplanes, and during campus internet outages.
Where a Closed API Would Have Been Better (Honest Trade-off): During factual spot-checks, the 1.5B open-weight model occasionally conflated action items containing dates with standalone deadline categories (e.g., filing a video demo deadline under action items instead of dates). A frontier closed model (like Claude 3.5 Sonnet or GPT-4o) has stronger semantic classification zero-shot out of the box. However, our Hallucination Guard mitigates this by anchoring every extracted item to its original line number regardless of category.

My Agent Session

The Lazy Payoff

To ensure complete empirical honesty, here are the measured numbers from my own daily routine:

Daily Activity	Without Sloth Digest (Manual Scrolling)	With Sloth Digest (Morning Glance)	Measured Daily Time Saved
Morning Class Council Catch-Up (180 messages)	14 minutes	45 seconds	13 min 15 s
Hackathon Team Chat Triage (75 messages)	9 minutes	30 seconds	8 min 30 s
Family Group Log Check (60 messages)	6 minutes	20 seconds	5 min 40 s
Total Daily Investment	29 minutes	1 minute 35 seconds	27 minutes 25 seconds

Net result: By being radically lazy and letting a local model read while I sleep, I save over 27 minutes every single morning without missing a single commitment or deadline.

Sleep peacefully. Wake up informed. Stay lazy. 🦥

Dhiraj201226

Closing the Freelance Skill Gap: From Client RFP to Winning Proposal in 2 Seconds

Dhiraj Shirse — Sun, 18 Oct 2026 22:15:30 +0000

*This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

JobMatch AI is an intelligent freelance gig and job matching platform that turns the tedious freelance job hunt into an unfair competitive advantage.

The Problem It Solves

Freelancers and side hustlers on platforms like Upwork, Fiverr, and Contra spend 30–45 minutes customizing bids for every individual client posting. Applying with generic proposals yields a sub-5% interview rate, while manually tailoring resumes, identifying missing requirements, and writing consultative proposals burns hours of unpaid time.

What JobMatch AI Does (In Under 2 Seconds)

0–100 Match Calibration: Evaluates candidate resumes or profiles against client job descriptions and provides a transparent, calibrated match score with 1–2 sentences of concise justification.
Skill Gap Diagnostics: Instantly extracts verified matched skills vs. missing client requirements.
3 Concrete Actionable Fixes: Generates three distinct, actionable suggestions to bridge gaps, improve positioning, or adjust the resume before submitting.
Tailored Proposal Generator: Drafts high-converting, personalized freelance proposals in the candidate’s chosen persona (Persuasive Freelancer, Technical & Analytical, or Consultative Partner).
Save & Compare Leaderboard (/saved): Allows freelancers to save multiple job postings to an interactive leaderboard, ranking them by match score so they only bid on the gigs they are mathematically most likely to win.
10-Second Judge Quick Try: One-click demo scenarios on the home screen to test full-stack developer and AI engineer scenarios with a single click.

Who Would Pay For It?

Active Freelancers & Agencies: Bidding on 10–30 client RFPs weekly. Saving 20 hours a month on proposal writing easily justifies a $9–$19/month subscription.
Side-Hustle Tech Seekers: Developers looking to monetize spare weekend hours by instantly finding client gigs that fit their exact stack without wading through thousands of listings.

Demo

Live Deployed App: https://job-match-v7ul.vercel.app (Replace with your live Vercel URL)
GitHub Repository: https://github.com/Dhiraj201226/Job-Match

Step 1: The Command Center & 10-Second Quick Try

The landing screen offers instant exploration. Judges and users can upload their own resume (PDF or pasted text) or click one of the 10-Second Judge Quick Try scenarios (Full-Stack React & Node MVP or Senior AI Application Engineer) to populate real-world candidate and RFP data with a single click.

Step 2: 0–100 Match Calibration & Skill Diagnostics

Once you click "Analyze Match & Write Proposal", Google Gemma 4 decomposes both documents, returning structured diagnostics in ~1.4 seconds:

Calibrated Match Fit: An objective 0–100 score (e.g., 82% Match Fit) with win probability classification.
Executive Justification: 1–2 sentences explaining why the candidate received the score based on tech stack alignment and experience.
Matched Skills Tags (Green): Verified skills satisfying client deliverables (e.g. React 19, TypeScript, Node.js, RESTful APIs).

Step 3: 3 Concrete Things to Fix & Tailored Proposal Generator

Rather than leaving candidates wondering how to improve, JobMatch AI provides three prioritized action items followed by a client-tailored proposal:

1. Concrete Resume Tweaks: Actionable bullets telling the candidate exactly what keywords or projects to add to pass screening filters.
2. Project Alignment: Suggestions on how to frame past project metrics to match the client's problem.

Step 4: Opportunity Ranking & Compare Leaderboard (`/saved`)

Freelancers shouldn't bid blindly. The Opportunity Ranking System allows candidates to save multiple client RFPs into a ranked pipeline:

Aggregates portfolio metrics: Total Saved, Top Match (e.g., 94%), Average Score, and Top Category.

Step 5: Transparent Side Hustle Economics

Unit economics make or break a side hustle. JobMatch AI features a transparent freemium pricing structure:

- Unlimited Free Local Mode: 100% on-device inference via local Ollama with zero token cost ($0.00).

Code

Dhiraj201226 / Job-Match

To check job matching skills

JobMatch AI — Freelance Gig & Job Matcher

DEV Hacktoberfest Open-Source AI Challenge · Week 2 "Side Hustle"
Powered by Google's open-weight model: Gemma 4 31B

JobMatch AI turns the tedious freelance job hunt into an unfair competitive advantage. Built around Google Gemma 4, it analyzes candidate resumes against job descriptions or freelance gig postings, calculates a precise match score (0–100), pinpoints missing skills, provides 3 concrete actionable fixes, and crafts tailored, high-converting proposals in the candidate's chosen tone.

📸 Product Demo & Walkthrough

Key Features

Match Score & Reasoning: 0–100 match calibration with concise justification explaining fit.
Skill Gap Diagnostics: Real-time extraction of matched skills vs. missing requirements.
3 Concrete Things to Fix: Actionable recommendations to bridge gaps before submitting.
Tailored Proposal Generator: Custom-tailored Upwork/freelance proposal written in your persona (e.g. Persuasive Freelancer, Technical, Consultative).
Save & Compare Ranking List: Save multiple gigs to a ranked leaderboard (…

View on GitHub

The repository is organized into a clean full-stack architecture:

frontend/: React 19 + Vite + Tailwind CSS with client-side PDF parsing and interactive Radar diagnostics.
frontend/lib/llm.js: Core open-model provider module handling prompt construction, schema validation, and retry mechanics.
api/ & frontend/api/: Vercel serverless edge functions (/api/analyze and /api/status) with zero cold-starts and 24/7 uptime.
frontend/test/llm.test.js: Automated unit test suite verifying structured JSON parsing, error recovery, and provider switching (8/8 tests passing).

How I Built It

JobMatch AI is built around Google's flagship open-weight model: Google Gemma 4 31B.

Architecture & Dual-Execution Engine

To ensure both lightning-fast public evaluation and zero-data-leak privacy, JobMatch AI features a single provider layer (frontend/lib/llm.js) supporting two switchable modes:

Cloud Mode (gemma4:31b-cloud):
- Deployed on Vercel's global edge network.
- Routes inference securely via serverless backend functions to Ollama Cloud.
- Evaluates complex 1,500+ token resumes and client RFPs in ~1.4 seconds.
- API keys are protected exclusively on the serverless backend (never exposed to the browser).
Local Mode (gemma4 on localhost:11434):
- 100% On-Device & Zero Data Leaks.
- All candidate data stays locally on your laptop without touching the internet.
- Automatically detects installed local Ollama models (gemma4:12b, qwen2.5:1.5b, etc.).

Robust JSON Schema Enforcement & Self-Correction

Small and mid-sized open models can occasionally produce conversational preamble or markdown code blocks. To guarantee 100% production reliability:

A custom extraction utility (extractJSON) cleans code fences and trailing commas.
A validator verifies schema keys (match_score, score_reason, matched_skills, missing_skills, 3 fixes, and cover_letter).
If validation fails, an automated one-time self-correction retry sends the exact validation error back to Gemma 4 to correct itself before rendering.

Why Does Open Innovation Matter?

In high-volume job seeking and freelance bidding, the model architecture is not an implementation detail — open models are what make this product economically and ethically viable:

1. Cost Per Match vs. The Proprietary API Tax

Commercial closed models (like GPT-4) charge recurring per-token fees for input context and output generation. A thorough evaluation requires sending a full candidate resume (~1,200 tokens) alongside a lengthy client RFP (~1,000 tokens), plus generating a structured proposal (~800 tokens). At scale, proprietary API fees destroy side-hustle unit economics.

With Gemma 4: In Local mode, the cost per match is $0.00. In Cloud mode, open weights allow deployment on cost-effective infrastructure with predictable unit economics.

2. Radical Candidate Data Privacy

Resumes contain highly confidential personal identifiable information: full legal names, contact details, employer history, salary markers, and proprietary project details.

With closed APIs, that data is transmitted to third-party proprietary servers. With Gemma 4's open weights, candidates can run evaluations completely on-device on their own hardware—guaranteeing 100% confidentiality.

3. Freedom from Vendor Lock-In & Portability

Proprietary models are black boxes subject to sudden deprecation, aggressive rate limiting, and unannounced behavioral shifts. Gemma 4's open weights allow developers to:

Quantize the model (GGUF 4-bit) for edge laptops.
Deploy dense 31B parameters on dedicated serverless clusters for maximum reasoning.
Fine-tune on winning freelance proposal corpora without third-party licensing restrictions.

Author: @Dhiraj201226

NatureQuest AI — Explore More, Scroll Less

Dhiraj Shirse — Sun, 11 Oct 2026 21:35:39 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

NatureQuest AI is a full-stack, gamified outdoor exploration platform built around a single mission: Explore More, Scroll Less.

Modern screen fatigue is real. Most fitness and wellness trackers only count steps or screen hours without giving people a compelling, adventurous reason to step outside and observe the real world.

NatureQuest AI changes that by turning your local environment into an interactive outdoor RPG:

1)Dynamic Local Micro-Quests: Procedurally generated nature quests categorized by difficulty, duration, and sensory focus (e.g., “Find tree bark moss facing north”, “Photograph an urban pollinator at work”, “Identify canopy leaf venation”).

2)Anti-Cheat AI Media Forensics (NeuralEye Engine): Anyone can download a photorealistic Midjourney, DALL-E, or Gemini image from their couch. NatureQuest uses an open-weight vision transformer ensemble to inspect submitted photos for real optical camera characteristics, rejecting synthetic AI diffusion, digital screen captures, and recycled images.

3)Gamified Outdoor Habits: Earn XP, level up through outdoor ranks (from Seedling Scout to Canopy Guardian), build consecutive outdoor streaks, and unlock nature badges.

4)AI Dynamic Quest Generation: Generates contextual outdoor exploration missions tailored to current weather, surroundings, and seasons using open-weight local LLMs.

Demo

hactoberfest-touch-grass.vercel.app

Code

Dhiraj201226 / hactoberfest-touch-grass

NatureQuest AI — Explore More, Scroll Less

A production-grade, gamified outdoor exploration platform that rewards users for spending time in nature and documenting authentic discoveries while combating screen addiction.

🌟 Core Product Vision & Philosophy

In an era saturated with endless algorithmic feeds and generative AI deepfakes, NatureQuest AI shifts focus from the virtual world to the living, physical world.

The Golden Rule:

Inspect your mission, put your phone in your pocket, and explore the living outdoors.

Users select real-world quests, step outside, record authentic photographs of their discoveries, and submit them to NeuralEye™—a multimodal media forensics pipeline that verifies camera sensor optical capture, detects diffusion artifacts, and enforces strict cryptographic duplicate protections.

🛠️ Tech Stack & Architecture

Frontend

Framework: React 19 + TypeScript + Vite
Styling: Tailwind CSS with custom editorial nature design tokens (forest-800, forest-900, lime, ivory)
Icons & Motion: Lucide Icons…

View on GitHub

How I Built It

Open-Weight AI Architecture NatureQuest is built entirely on open-source, open-weight AI rather than proprietary black-box APIs:

NeuralEye Multi-Layer Forensics (haywoodsloan/ai-image-detector-deploy):

Architecture: SwinV2 (Hierarchical Shifted-Window Vision Transformer) fine-tuned for high-accuracy discrimination between physical sensor photons and synthetic diffusion artifacts.
Ensemble Defense:
Layer 1 (Fast-Path Provenance): Inspects binary byte arrays for C2PA cryptographic manifests (Google Imagen, Adobe Firefly), imperceptible SynthID watermarks, and visible 4-point generative corner stars in under 5ms.
Layer 2 (Vision Transformer): Evaluates frequency domain artifacts and lighting consistency using SwinV2 running on CPU inference.
Anti-Duplicate Hashing: Uses perceptual difference hashing (dhash) and SHA-256 to prevent users from recycling previous quest photos.
Procedural Quest Generation (Qwen/Qwen2.5-1.5B-Instruct via Ollama):

An open-weight Apache 2.0 licensed SLM (Small Language Model) generates micro-quests with structured JSON outputs specifying title, category, sensory observation guidelines, difficulty, and XP rewards.
Runs locally or on edge containers without per-token API costs or third-party data tracking.

High-Performance Engineering Highlights Client-Side Image Optimization: Smartphone cameras produce 10MB+ raw files. We built an in-browser canvas compressor that scales photos to 1600px and 82% JPEG quality before transmission. This dropped upload payload sizes by 95% (from 10MB down to ~300KB), reducing verification latency from 30s to under 1s. Background Pipeline Pre-warming: The SwinV2 PyTorch weights are pre-warmed in a background thread on application boot so end-users never experience cold-start latency.

Why Does Open Innovation Matter?

Open innovation is the backbone of NatureQuest AI for three critical reasons:
1)Privacy for Physical Locations & Personal Captures: Asking users to upload personal outdoor photos to proprietary closed-source APIs means their real-world routines, geolocation EXIF data, and surroundings could be used to train closed models. Open-weight models deployed on self-hosted infrastructure guarantee user privacy.
2)Deterministic, Inspectable Forensics: A verification system must be transparent. If an image is rejected, open-source weights and inspectable C2PA/SynthID heuristics allow us to explain why an image failed (e.g. generative artifacts vs. resolution issues), rather than hiding behind an opaque corporate score.
3)Zero Paywalls & Accessibility: Health and outdoor wellness tools should not be tied to expensive API subscription tiers that shut down when pricing doubles. Open-weight models like Qwen 2.5 and open Hugging Face vision transformers ensure anyone can run and self-host NatureQuest anywhere.

Prize Categories

Render — Best Use of Render**

NeuralEye uses AI-powered image detection to distinguish AI-generated images from authentic photographs, supporting our nature-exploration platform. We plan to use Render to deploy and host the backend inference API, enabling reliable access to the AI detection model for image verification.

DEV Community: Dhiraj Shirse

Sleep Peacefully. Wake Up Informed: How I Built Sloth Digest with Local AI

What I Built

Demo

Code

Dhiraj201226 / Sloth-Digest

🦥 Sloth Digest

The Problem

The Solution: Radical Laziness

Tech Stack Highlights:

How I Built It

Why Does Open Innovation Matter?

My Agent Session

The Lazy Payoff

Closing the Freelance Skill Gap: From Client RFP to Winning Proposal in 2 Seconds

What I Built

The Problem It Solves

What JobMatch AI Does (In Under 2 Seconds)

Who Would Pay For It?

Demo

Step 1: The Command Center & 10-Second Quick Try

Step 2: 0–100 Match Calibration & Skill Diagnostics

Step 3: 3 Concrete Things to Fix & Tailored Proposal Generator

Step 4: Opportunity Ranking & Compare Leaderboard (/saved)

Step 5: Transparent Side Hustle Economics

- Unlimited Free Local Mode: 100% on-device inference via local Ollama with zero token cost ($0.00).

Code

Dhiraj201226 / Job-Match

To check job matching skills

JobMatch AI — Freelance Gig & Job Matcher

📸 Product Demo & Walkthrough

Key Features

How I Built It

Architecture & Dual-Execution Engine

Robust JSON Schema Enforcement & Self-Correction

Why Does Open Innovation Matter?

1. Cost Per Match vs. The Proprietary API Tax

2. Radical Candidate Data Privacy

3. Freedom from Vendor Lock-In & Portability

NatureQuest AI — Explore More, Scroll Less

What I Built

Demo

Code

Dhiraj201226 / hactoberfest-touch-grass

NatureQuest AI — Explore More, Scroll Less

🌟 Core Product Vision & Philosophy

🛠️ Tech Stack & Architecture

Frontend

How I Built It

Why Does Open Innovation Matter?

Prize Categories

Render — Best Use of Render**

Step 4: Opportunity Ranking & Compare Leaderboard (`/saved`)