Skip to content
View somil71's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report somil71

Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
somil71/README.md

Somil Jha

Systems builder. I write Rust, Python, and TypeScript — and I ship things that work.

somilwork77@gmail.com · LinkedIn · Instagram


What I've shipped

Needle — Local-first code search engine

Built entirely in Rust from scratch. Hybrid BM25 + HNSW vector search, hand-rolled (no Qdrant, no Tantivy). Tree-sitter AST chunking across 8 languages, live D3 call graph with community detection, MCP server with 11 tools. Ships as a native Tauri desktop app, VS Code extension, CLI, and cloud deployment with GitHub OAuth.

Rust Axum Tauri Tree-sitter HNSW BM25 MCP


GenMol — AI Molecular Generation Platform

Full replication of Lee et al. 2025 (arXiv:2501.06158) — a Masked Diffusion Language Model over SMILES sequences with transformer-encoder denoiser, monotonic noise schedule, and iterative demasking inference. Includes a 3-layer hallucination guard (RDKit + perplexity + sliding window) with automatic checkpoint rollback, and a nightly self-improvement loop that retrains on high-quality production molecules.

Deployed as a microservices stack: React + Firebase Auth frontend, Node.js/Express gateway, Flask/PyTorch ML service, Supabase (PostgreSQL) data layer, Kubernetes HPA for auto-scaling.

PyTorch Flask React Kubernetes Supabase RDKit GitHub Actions


FairPay — Decentralized Freelancer Escrow Protocol

Production-grade multi-milestone escrow on EVM (Sepolia). Funds lock at contract creation; milestones release via signatures, GitHub commit hashes, or IPFS CIDs — with automatic 48-hour default release if the client goes silent. Partial disputes let clients contest exactly the portion they disagree with. Soulbound reputation NFTs on completion; stake-based arbitration to punish bad-faith disputes.

Solidity Next.js wagmi Prisma Neon Postgres BullMQ Docker Kubernetes


Job-Search Multi-Agent System

14-agent autonomous pipeline: discovers jobs across 15+ boards, scores fit against your resume, tailors the resume (A3) and validates it (A4) in a bounded feedback loop, generates a per-job apply kit, monitors your Gmail inbox for replies, classifies emails (rejection / interview / assessment / offer), schedules .ics calendar invites, tracks assessment deadlines, sends follow-ups after 3-4 days of silence, and produces weekly diagnostics reports. Claude optional; runs fully offline by default.

Python FastAPI Playwright Gmail API SQLite Claude API


BSES Delhi AI Chatbot — Multilingual Electricity Service Assistant

Production-ready multi-tenant chatbot replacing a Yellow.ai subscription. A scikit-learn TF-IDF + LinearSVC classifier trained on ~36,000 labelled examples handles English, Hindi (Devanagari), and Hinglish (Latin-script) in a single model — ~1ms inference, no GPU. Three routing layers: L1 (BSES API calls), L2 (static knowledge base), L3 (Groq Llama 3.1 synthesis). Gemini Vision reads uploaded bill photos and extracts structured data. Redis caching, JWT multi-tenant sessions, per-site rate limiting. 387 tests, all passing.

FastAPI scikit-learn Redis SQLite Groq Gemini Docker


Decision Intelligence Platform

End-to-end decision capture and drift monitoring. Slack-style chat UI where teams log decisions; a FastAPI AI microservice (GPT-4o-mini) extracts structured data and auto-generates tasks. A cron job re-evaluates decisions for drift and flags at-risk items. Circuit-breaker + retry logic throughout.

React Express FastAPI MongoDB Redis Docker


FITBOARD — Shoppable Editorial Lookbook

Dark-luxury fashion platform where every outfit is a clickable product map. Admins pin individual garments on photos with exact pixel coordinates; each pin links to a tracked affiliate redirect. Infinite-scroll masonry feed, Framer Motion spring-physics pin reveals, SVG connector lines from cards to hotspots, Supabase RLS, Upstash Redis rate-limiting on click events.

Next.js Supabase Framer Motion Zustand React Query TanStack


AI Consistency Auditor

Rule-based data auditing system — upload two JSON snapshots, define rules (NON_DECREASING, IMMUTABLE, NON_NULL), run an audit to detect violations with severity classification and plain-English explanations. No ML — deterministic rule engine, full audit trail, Docker deployment.

FastAPI React MongoDB Docker


Stack

Languages

Rust Python TypeScript JavaScript Solidity C++ Java

Frontend

React Next.js Tailwind

Backend & Systems

Axum FastAPI Node.js Express

AI / ML

PyTorch scikit-learn

Infrastructure

Docker Kubernetes AWS Redis


Stats


Open to client work · somilwork77@gmail.com

Pinned Loading

  1. AI-auditor AI-auditor Public

    Python

  2. AI_VAULT AI_VAULT Public

    TypeScript

  3. RESUME_ANALYSIS RESUME_ANALYSIS Public

    Python

  4. INNOVEDA INNOVEDA Public

    JavaScript