DEV Community: Priyanshu Bharti

AI Buddies: Bringing Digital Ghosts to Life with Open-Weight Models

Priyanshu Bharti — Tue, 27 Oct 2026 08:11:03 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

What I Built

Ghost in the Machine: AI Buddies as a Digital Ofrenda is a custom feature extension built inside the AI Buddies platform that functions as a Virtual Ofrenda & Memory Guardian.
Combining the reflective spirit of Día de los Muertos with the spooky atmosphere of Halloween, this tool ensures that the early code, contributions, and lessons of open-source friends, mentors, and forgotten past projects do not fade away. It channels open-weight models to act as a "ghost in the machine"—whispering back memories, preserving ancestral or early coding struggles, and telling stories of how past open-source tools paved the way for modern development.

How I Built It

To capture an eerie, offline, and deeply reflective atmosphere, the architecture relies entirely on open-source AI foundations:
• Frontend Architecture: Built using React, TypeScript, and Tailwind CSS. Custom styling invokes a glowing, candle-lit digital ofrenda aesthetic with full markdown and KaTeX support.
• Open AI & Agent Frameworks: Integrated local inference wrappers, LangChain agent harnesses, and Hugging Face open-weight models. This allows the memory-retrieval agent to run complex local narrative generation without hitting external servers.
• Offline Data Vault: Uses client-side storage paired with local model weights, ensuring personal stories, family oral histories, or remembrance notes never leave your personal machine.

Why Does Open Innovation Matter?

When dealing with personal memories, family archives, and tributes to loved ones, open innovation provides three non-negotiable advantages over closed APIs:
• Absolute Data Privacy: Commercial APIs process intimate family histories and personal memories on corporate servers. Open-weight models running locally keep all sentiment and personal records completely private.
• Offline Permanence: Día de los Muertos honors what endures. Closed APIs can deprecate endpoints, change terms of service, or shut down overnight. An open-source approach means this memorial tool can run on a laptop with zero internet connection, preserving memories indefinitely.
• Custom "Spooky" Tuning: Closed chat models enforce strict, neutral corporate personas. Open-weight models allow us to rewrite system prompts completely, letting the agent channel historical developer lore and respectful remembrance tones authentically.

My Agent Session

The Spirit of It

When testing the virtual memory ofrenda with a fellow developer, we fed in notes about an old, abandoned open-source project we worked on years ago. The local model generated a hauntingly accurate recap of our old debugging nights, complete with nostalgic styling. It felt like bringing a digital ghost back to life—reminding us that behind every repository lies human effort, memory, and community.

SubScribe Field OS: A Local-First Voice OS for Construction Paperwork

Priyanshu Bharti — Mon, 19 Oct 2026 20:38:17 +0000

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

What I Built

SubScribe Field OS is a local-first, voice-controlled operating system engineered to automate construction field documentation entirely offline. On loud, chaotic job sites, tradespeople can speak a single sentence and instantly generate a signed, print-ready change order, daily log, or delay notice.
Small trade subcontractors (like electricians, plumbers, and HVAC technicians) lose thousands of dollars annually because out-of-scope work rarely gets documented on time. They are busy working on ladders, not sitting at desks. SubScribe allows them to translate plain spoken language—such as "We moved four electrical boxes eight feet due to a duct obstruction"—into a structured, contract-linked PDF in seconds.
Who benefits: Small trade contractors who need to safeguard their profit margins, eliminate manual paperwork, and secure faster payments for extra work. Recovering just one unbilled change order easily covers the system's value. The repository bundles the complete local stack: audio transcription, schema routing, contract RAG, a local SQLite audit trail, and a PDF generation engine.

How I Built It

The entire application runs locally on a standard laptop with zero API keys or per-document processing fees.
• Local Audio Transcription (faster-whisper): Audio is processed using an int8 quantized version of faster-whisper running directly on the local CPU. To overcome heavy job site background noise, a targeted trade-vocabulary prompt (loaded with terms like conduit, junction box, HVAC, rough-in, EMT) is injected into the model, ensuring highly accurate technical transcriptions.
• Schema-Enforced Routing (Gemma 2B + Ollama + Pydantic): Raw text transcripts are routed directly to Gemma 2B via Ollama using temperature=0.0 and format="json". Strict Pydantic schemas (ChangeOrderOutput, DailyLogOutput, DelayNoticeOutput, and an IntentClassification router) validate every model response. If a format mismatch occurs, the orchestrator instantly catches it and retries, creating an enterprise-grade financial tool.
• Offline Contract RAG (ChromaDB): Before finalizing a change order, the system searches a local ChromaDB instance populated from the subcontractor's original scope agreement PDF. The most relevant clauses are provided to Gemma as context, enabling the tool to automatically cite the exact contract sections backing up the financial claim.
• Local Audit Trail (SQLite): All finalized records are serialized via Pydantic and written to a local SQLite database across three dedicated tables (daily_logs, change_orders, delay_notices), bypassing external syncing or database conflicts.
• Printable PDF Engine (Jinja2 + WeasyPrint): A HTML template dynamically populates a professional invoice layout including company branding, out-of-scope descriptions, contract clause citations, material/labor cost breakdowns, and a physical signature block for the project supervisor. WeasyPrint compiles this into a clean PDF in under two seconds.
• Field UI (Streamlit/Gradio & FastAPI): Field users interact with a lightweight interface to record audio or upload .wav files, review the parsed JSON, and export PDFs. Off-site managers can track history via a FastAPI review dashboard.

Why Does Open Innovation Matter?

Open innovation is a mandatory requirement for this business model to protect users and control operational costs:

Privacy is the Product: Change orders contain sensitive labor rates, material margins, and confidential legal agreements. Uploading this data to third-party cloud APIs presents significant liabilities. Running Gemma 2B locally guarantees that proprietary contract data, audio files, and financials never leave the contractor’s physical hardware.
Zero Marginal Costs: Closed APIs levy charges per token. A contractor handling dozens of records daily would see usage fees cut into their profits. Local inference keeps runtime costs at zero, making the software highly affordable for solo operators.
True Offline Independence: Job sites and basements frequently lack reliable internet connections. A cloud-reliant tool breaks down right when a contractor needs to document an obstruction before their crew leaves the area. Local open weights ensure absolute operational reliability anywhere.
No Vendor Lock-In: Open-source architecture permits seamless model swapping. If a stronger open-weights model is released, or an older laptop demands a lighter footprint, the core system can adapt without breaking the application logic.

The Lazy Payoff

The system was put to the test by a solo contractor running SubScribe directly on a laptop. By simply dictating daily logs and change orders aloud, he was able to instantly compile polished, legal-ready PDFs without touching a keyboard. The immediate payoff convinced him that he would readily pay a $5–$10/month subscription—proving that a zero-marginal-cost architecture unlocks a highly competitive, accessible pricing model for trade professionals.

Prize Categories

• Best Use of Gemma — Gemma 2B serves as the local intelligence foundation, managing all intent classification, schema-enforced JSON extraction, and contract clause referencing via Ollama.
• Best Use of Render — The FastAPI review dashboard is hosted on Render's free tier, allowing office managers to view synchronized data logs seamlessly as soon as field devices reconnect to Wi-Fi.
To help fine-tune this submission, please let me know:
• Do you want to include any specific GitHub repository links or live deployment URLs?
• Would you like to add details about any additional open-weights models you might have tested alongside Gemma 2B?
• Should we expand the Lazy Payoff section with more concrete metrics from your friend's test run?

PhishBrief AI: A Privacy-First, Local AI Phishing Assistant for Small Businesses

Priyanshu Bharti — Tue, 13 Oct 2026 16:47:43 +0000

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

What I Built

PhishBrief AI is an intelligent, locally-hosted phishing analysis assistant tailored for individuals and small businesses.

Cybercriminals constantly use urgency, spoofing, and malicious links to steal sensitive data, yet many freelancers and small teams simply cannot afford expensive enterprise email security platforms. PhishBrief AI bridges this gap by acting as an accessible first line of defense. It evaluates suspicious emails, highlights red flags, and offers actionable security advice—all without transmitting your data to the cloud.

As a side hustle, this project represents a lightweight, affordable security triage solution designed specifically for IT support teams, students, and small organizations that need understandable threat analysis without the enterprise price tag.

How I Built It

I engineered PhishBrief AI by combining standard web frameworks with powerful local AI inference.

Application & UI: The core logic is written in Python, utilizing Streamlit to create an interactive, user-friendly interface.
Local Inference Engine: I used Ollama as the local runtime to ensure all data processing happens directly on the user's machine.
Open-Weight Models: The primary engine driving the analysis is Granite 3.2 Vision, carefully configured for this specific security workflow. The development environment also supports Gemma 3 1B as an alternative local model.

The architecture centers around a privacy-conscious workflow. When a user pastes a suspicious message, the local model parses the content, identifies phishing indicators, explains the exact risks in plain English, and generates safety recommendations. By using local inference rather than third-party APIs, the application ensures that potentially sensitive email data never leaves the device.

Why Does Open Innovation Matter?

Open-source AI is the foundational pillar of this project. Without open-weight models and runtimes like Ollama, building a truly privacy-first tool would require relying on proprietary, paid APIs—which introduces both data privacy risks and recurring costs.

Open innovation breaks down these barriers. It allows independent developers to experiment freely, iterate on prompting strategies, and build practical applications that respect user privacy. While open models don't automatically guarantee perfect accuracy, they give creators complete control over how the AI is deployed, evaluated, and improved without enterprise-level infrastructure.

My First User

My ideal target audience consists of freelancers and small-business teams who need a straightforward way to evaluate threats without requiring a background in cybersecurity.

I am currently onboarding early testers to run the tool against safe, simulated phishing emails. My primary objective at this stage is to gather feedback on whether the AI's risk breakdowns are intuitive and genuinely helpful for non-technical users. This validation will help refine the tool from a prototype into a viable side-hustle product.

Prize Categories

Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

OutBound AI: The Open-Source Recommender That Powers Offline Living

Priyanshu Bharti — Wed, 07 Oct 2026 20:53:04 +0000

What I Built

OutBound AI is an open-source intelligence engine designed to reverse the typical digital engagement loop. Instead of engineering software to maximize screen time, this app acts as a digital launch pad to the physical world. It is built for developers, students, and professionals who find themselves spending entire days trapped behind desks and monitors.
Users simply input their available time window, preferred activity intensity, and current mood. The system processes this context to deliver a single, friction-free outdoor objective. By providing a concrete, actionable plan rather than an endless scroll of choices, it removes the decision fatigue that keeps people glued to their chairs—guiding them out the door to walk, exercise, or rest.

How I Built It

The application architecture is built entirely around localized, open-weight AI. Rather than relying on cloud-hosted endpoints or treating the AI as an open-ended conversational chatbot, the system utilizes the model as a deterministic decision-making router.

AI Engine & Inference: Powered by a local open-weight large language model (LLM) running offline via local inference. The model evaluates user constraints (time, vibe, and energy) to output highly tailored outdoor blueprints.
Architecture: Built on a modern web stack for the user interface and application logic. The design decouples the application framework from the intelligence layer, ensuring that the underlying open model can be easily swapped out or updated as open-source hardware and software evolve.

Why Does Open Innovation Matter?

Open innovation was the vital catalyst for this project. Utilizing open-weight models grants developers true autonomy, completely bypassing restrictive, metered APIs and proprietary pricing structures. Building this locally provided absolute control over prompt alignment, system constraints, and data privacy.
More importantly, it allowed me to challenge the status quo of modern AI. AI is traditionally weaponized to capture user attention spans and optimize screen engagement. Open-weight innovation allowed me to explore the exact antithesis: utilizing locally controlled intelligence to systematically disconnect users from their machines.

My Agent Session

Throughout the development lifecycle, I collaborated with an AI pair-programmer to accelerate building the application. The agent served as a core utility—assisting with local environment configuration, resolving edge-case bugs in the local inference pipeline, and rapidly prototyping UI components.

Prize Categories

Hacktoberfest Open-Source AI Challenge: Touch Grass
Open-Source AI

FocusFlow: A Minimalist Deep Work Dashboard Built for a Friend

Priyanshu Bharti — Sun, 04 Oct 2026 09:59:35 +0000

What I Built

I built FocusFlow, a premium, distraction-free productivity dashboard engineered to trigger and sustain the "Flow State."

I originally built this for my friend, who struggles heavily with ADHD and context-switching fatigue. They were constantly getting distracted because their workflow required juggling four different apps: a Pomodoro timer, a Spotify window, a Trello Kanban board, and a day planner. Every time they switched tabs to change a song or check a task, they lost their focus.

FocusFlow solves this by consolidating everything into a single, sleek, horizontal interface. It features a high-precision Pomodoro timer, a unified visual time blocker, a minimalist Kanban board, and an integrated ambient sound engine. It even has a "Zen Mode" for pure, zero-distraction focus. By keeping everything in one hardware-accelerated dashboard, it completely eliminates app-switching friction.

Demo

🚀 Live Preview: focusflow-pb.vercel.app

Code

PriyanshuBh / focusflow

FocusFlow is an all-in-one productivity application designed to help you get into the "flow state" and stay there.

FocusFlow | Minimalist Deep Work Dashboard

A Sleek, Minimalist Sanctuary for Deep Work: Horizontal Pomodoro + Unified Time Blocker + Kanban + Ambient Sound Engine
View Demo »

View Code · Report Bug · Request Feature

Table of Contents

🤖 About The Project

FocusFlow is a premium, distraction-free productivity dashboard engineered to trigger and sustain the "Flow State." By consolidating a High-Precision Pomodoro Timer, a Unified Visual Time Blocker & Task Bucket Grid, a Minimalist Kanban Board, and an Integrated Ambient Audio Engine into a sleek, horizontal, hardware-accelerated interface, FocusFlow eliminates app-switching friction and cognitive fatigue.

Built with Next.js 15, React 19, PWA capabilities, and the Midnight Slate design system, it delivers a liquid-smooth 60fps experience even on modest hardware.

🔥 Key Features

🍅 Precision Pomodoro System &…

View on GitHub

How I Built It

FocusFlow is built with performance and aesthetics in mind. Here is the core stack:

Framework: Next.js 15 & React 19 (App Router)
Styling: Tailwind CSS 4 with a custom "Midnight Slate" design system
Animations: Framer Motion for GPU-accelerated layout transitions and spring physics
Audio: Web Audio API for native ambient soundscapes (Brown noise, Rain, Forest) & Spotify Web API
Architecture: Full PWA manifest & Service Worker for offline caching.

The AI Advantage:

To build this quickly over the weekend, I relied on open-source AI agent tooling to help architect the more complex parts of the application. Specifically, I used Antigravity to help me rapidly prototype the IntersectionObserver logic for the floating mini-widget timer and to write the boilerplate for the Web Audio API synthesizer. Having an AI agent act as a pair programmer allowed me to focus heavily on the UX and Framer Motion animations while the agent handled the boilerplate audio routing and state calculations.

Why Does Open Innovation Matter?

Open innovation is the only reason a solo developer can build a premium, 60fps productivity application over a single weekend.

Without open-source frameworks like Next.js, accessible primitives like Shadcn UI (Radix), and the vibrant open-source AI community providing accessible coding agents, building the smooth drag-and-drop Kanban or the zero-latency ambient audio engine would have taken weeks. Open innovation democratizes the ability to create "premium-feeling" software, allowing us to solve highly specific, personal problems for our loved ones without needing a massive corporate budget.

My Agent Session

Embedded session coming soon / Link to DevRelay session

Prize Categories

Build for a Friend
Best UI/UX
Productivity / Deep Work

Team Submissions: Built solo by Priyanshu Bharti