RAG pipelines with hybrid search, multi-agent orchestration (6 agents), cross-encoder reranking, evaluation harnesses

React 19 + TanStack Start frontends, FastAPI async backends, Clerk auth, enterprise RBAC

PostgreSQL + pgvector, Redis HNSW indexes, Docker, $0/month free-tier deployments

Open Source Contributor · OmniRoute — Remote Diagnosed an HTTP 400 regression in the memory-injection pipeline affecting strict LLM providers (Xiaomi MiMo); proposed a declarative systemMessageMustBeFirst Zod schema flag adopted by the maintainer into the shipped fix — 25/25 Vitest + 30/30 Node test-runner coverage (PR #6225). Built an accessible "Configured Only" filter toggle mapping live connection state to a filterable data grid — 168 additions across 4 files, 9/9 tests passing, shipped into v3.8.45 (PR #6245). Integrated Claude 5 Sonnet into the provider registry with regression test coverage, shipping a verified signed commit within hours of the model's release (PR #6209, v3.8.45). Audited and normalized 9 core docs + 20+ localized READMEs across 42 locales, zero regressions (PR #6105, v3.8.44).

Full Stack Developer Intern · nTheta Works Pvt. Ltd. — Remote Built a two-stage semantic retrieval pipeline (Ollama embeddings → Redis HNSW → FlashRank cross-encoder reranking) for NLPForge, an enterprise LLM API testing platform — improved template matching accuracy by 40% and reduced manual QA effort by 60%. Shipped async FastAPI microservices + Next.js/TypeScript dashboards with Docker Compose orchestration and CI/CD on Linux servers.

AI & Full Stack Developer · Freelance — Remote Built SmartLearnX, an AI-powered LMS, with a dropout prediction model (Logistic Regression, 91.4% accuracy) and academic performance forecasting (Random Forest, R² = 0.89) deployed as a FastAPI microservice alongside a React/Node.js frontend. Integrated NLP features (BERT for quiz generation, spaCy chatbot) serving 24/7 student support with sub-2-second response times under load.

Full Stack Developer Intern · OBG Outsourcing Pvt. Ltd. — Jaipur Led FinSageAI360, a multi-tenant financial intelligence SaaS — cut monthly close reporting time by 45% and reduced manual operational effort by 30% through AI-driven anomaly detection and real-time KPI dashboards. Designed a JWT-authenticated REST API (Node.js/Express/MongoDB) with granular RBAC for multi-tenant data isolation.

Software Developer Intern · Om Logistics Ltd. — Delhi Optimized enterprise document search by implementing LangChain + FAISS vector embeddings — reduced query latency by 70% across 10,000+ documents and improved retrieval accuracy by 40%. Built RESTful APIs (Node.js) to automate logistics workflows, eliminating 20% of manual data-entry tasks.

SIH 2023 National Winner — recognized by Coal India Limited & CMPDI

AI document intelligence platform for the Ministry of Coal. 6 specialized AI agents across 4 providers (Groq, Cerebras, Mistral, Gemini) running concurrently via asyncio. Production RAG with hybrid search (pgvector + BM25) + cross-encoder reranking, at $0/month infrastructure cost.

Next.js 16 React 19 FastAPI pgvector Supabase Upstash Redis Clerk Docker

🔗 GitHub · 🌐 Live Demo

Enterprise AI NLP platform

Two-stage retrieval pipeline: Ollama embeddings → Redis HNSW KNN search → FlashRank cross-encoder reranking. 8 LLM providers, 15+ embedding models. 40% accuracy improvement, 60% QA effort reduction.

Next.js 16 FastAPI SQLAlchemy 2.0 PostgreSQL 15 Redis Stack FlashRank Ollama

🔗 GitHub

AI financial intelligence platform

Multi-tenant SaaS for accounting feed ingestion, normalization, and AI-generated cashflow/risk/anomaly insights. Built at OBG Outsourcing. +45% faster reporting, +30% operational efficiency.

Next.js Node.js MongoDB Prisma JWT AI Analytics

AI-powered Learning Management System

Adaptive learning platform with dropout prediction (91.4% accuracy), performance forecasting (R² = 0.89), and 24/7 NLP chatbot support. Built as a freelance project.

React TypeScript Node.js FastAPI MongoDB Redis Docker ML

Shipped an accessible "Configured Only" filter for the provider-rankings dashboard — maps live /api/providers connection state to a filterable data grid with a "Status" column, role="switch" with aria-checked for screen reader accessibility, useEffect cleanup for memory-leak prevention, and a full Vitest suite. 168 additions across 4 files, 9/9 tests passing (PR #6245, v3.8.45).

Integrated Claude 5 Sonnet into the claude_web provider registry with a registry regression test, shipping a verified signed commit within hours of the model's release. PR #6209, v3.8.45.

Audited 9 core docs and ~20 localized READMEs across 42 locales, removing untranslated Portuguese/Chinese prose and correcting stale architecture facts (routing strategies 13→17, service modules 36→134). Passed docs-sync-strict CI gate — zero regressions (PR #6105, v3.8.44).

Diagnosed an HTTP 400 regression for strict LLM providers (Xiaomi MiMo) caused by system-message ordering in the memory-injection pipeline. Proposed a declarative Zod schema with a systemMessageMustBeFirst flag — adopted by the maintainer into the broader shipped fix. 25/25 Vitest + 30/30 Node test-runner coverage (PR #6225).