| 5.0★ Freelance client rating (Upwork) |
Live AskMyDocs eval dashboard, CI-gated |
1 Model published on Hugging Face |
Final-year B.Tech IT student building AI systems that ship — full architecture, backend, frontend, and CI, deployed on Render / Vercel / Streamlit. My flagship work: AskMyDocs, a RAG document Q&A tool that cites its exact source and refuses to guess when the answer isn't there. And SalesAgent, a lead-scoring agent that turns a LinkedIn URL into a researched cold email in under a minute.
- 🎓 B.Tech IT, MITS Gwalior (Final Year, 2023–2027)
- 🔭 Specializing in RAG pipelines, multi-agent systems, LLM fine-tuning (LoRA/PEFT), and Model Context Protocol (MCP) tool serving
- 🏆 NPTEL (IIT Kanpur) — Elite, Top 5%, Cloud Computing & Distributed Systems (90%)
- 💼 Freelance AI Developer on Upwork (5.0/5.0 client rating)
1. AskMyDocs — RAG Document Q&AAnswers questions over 50-page PDFs in under 3 seconds, returning the exact source chunk and cosine similarity score behind every answer — and withholds an answer instead of hallucinating when the document doesn't cover it. Ships with an LLM-as-judge + keyword-validation eval pipeline wired into CI to catch retrieval regressions before deploy. Current eval numbers are always on the live dashboard, not hardcoded here. 
|
CI-gated retrieval eval, not just a demo. |
2. SalesAgent — Autonomous B2B Sales AgentPaste a LinkedIn URL — a LangGraph research node pulls real signal, a Random Forest model scores the lead, and Groq drafts a hyper-personalized cold email referencing actual company events (94/100 on the Satya Nadella demo run, referencing a real, live job posting). End to end in under 45 seconds. A self-built eval harness caught two production bugs before they shipped — uniform lead scores and a missing sender identity in generated emails. 
|
Full agentic loop + ML scoring + eval harness. |
3. LoRA Fine-Tuned Resume Screener — Published on Hugging FaceFine-tuned a LoRA adapter (r=16, just 0.44% of parameters trained) on Qwen2.5-0.5B so structured JSON resume-fit verdicts are the model's default output — not something coaxed out with prompting. Benchmarked against zero-shot on the same 96-example eval set: verdict accuracy went from 17.7% to 88.5%, and mean score error dropped from 32.03 to 5.39 points — fine-tuning fixed judgment, not just output format. Validation loss tracked training loss across 3 epochs with no divergence, confirming no overfitting. The hosted demo serves scoring through a Groq-hosted backend for free-tier hosting reasons; the trained adapter itself runs locally with no API key (details in the repo's Deployment Note).
|
Training work, not just inference — and published. |
4. AgentLoop — Multi-Step Research AgentNot a chatbot — a research agent that decomposes a question into sub-questions, searches the live web, reflects on gaps in its own notes, loops back, and delivers a fully cited report. Two-tier memory (short-term run state + long-term SQLite recall) streams live trace events to the UI as it reasons.
|
Plan → act → reflect → loop, with visible reasoning traces. |
5. Self-Healing RAG — Critique-and-Retry RAG PipelineRAG pipeline that grades its own answers against the retrieved context — if a response isn't grounded, it reformulates the query and retries instead of returning an ungrounded answer.
|
Self-critique loop, not a one-shot retrieval. |
▶ More projects (voice AI, automation agents, MCP server & more)
LLM Cost Router — Live Demo
Heuristic query-complexity classifier that routes requests between a cheap and a far more expensive Groq model, cutting cost significantly on simple queries with no quality loss on complex ones. Live dashboard tracks real spend vs. a same-model baseline.
FastAPI Groq (Llama 3.1 8B / 3.3 70B) Streamlit
AI Interview Coach — Live Demo
Real-time voice interview simulator — answers scored on relevance, clarity, technical accuracy, and confidence via Groq LLaMA 3.3, with a downloadable PDF report.
Streamlit Faster-Whisper Groq PDF Generation
Agentic RAG Research Assistant — Live Demo | API Docs
LangGraph tool-routing RAG system — retrieves grounded answers from uploaded PDFs via Chroma, declines out-of-scope questions, and routes queries between a cheap and large model based on complexity.
LangGraph FastAPI Streamlit Chroma Groq
AI Data Analyst Agent — Live Demo
Upload CSV, Excel, PDF, Parquet, XML, SQLite, ODS, or Feather files — ask questions in plain English, get instant charts and insights.
Streamlit Groq pandas
Email Agent — Live Demo
AI Gmail agent that classifies emails and drafts context-aware replies you can approve or edit before sending.
IMAP SMTP Groq LLaMA 3.3 Streamlit
ARIA – Voice AI Assistant — Live Demo
Speech-to-speech AI assistant with 99-language support and conversation memory. Speak in any language — ARIA transcribes, thinks, and talks back.
FastAPI Faster-Whisper Groq LLaMA gTTS
ResumeIQ — Live Demo
AI resume screener that scores ATS compatibility, identifies gaps, and exports detailed PDF reports.
Python Flask Groq
StartupScope — Live Demo
Multi-agent CrewAI crew — Researcher, Analyst, and Writer agents collaborate to search the web and generate structured startup intelligence reports.
CrewAI Groq Streamlit
JobHunt — Live Demo
AI-powered Telegram job aggregator — scores every post and auto-applies via email or form-fill. Watches job channels 24/7 so you don't have to.
FastAPI PostgreSQL Groq
Portfolio MCP Server — Published on PyPI & the MCP registry
A working MCP server exposing 5 tools — project search, stack filtering, resume summary — so any MCP client queries this portfolio as live, structured, callable data instead of a static page. pip install portfolio-mcp-server and it's live in any MCP client in under a minute.
Python MCP (FastMCP SDK) stdio transport Claude Desktop
North Star Support Chatbot — Local only
AI-powered customer support chatbot for a North Star outdoor gear store, with full conversation handling and escalation logic.
React FastAPI Groq LLaMA
▶ Archived
n8n Email → Slack — Archived, hosting suspended
No-code AI automation pipeline: fetches unread Gmail → summarizes with Groq LLaMA → detects priority → pushes digest to Slack.
n8n Groq Gmail Slack
Freelance AI Developer — Self-Employed, Remote · May 2026 – Present Design and ship deployed AI systems end-to-end for clients — agentic workflows, RAG pipelines, LLM tooling. Delivered a chatbot contract via Upwork Talent Accelerator end-to-end in 3 days (5.0/5.0 client rating).
Cloud Computing and Distributed Systems — NPTEL (IIT Kanpur) Elite + Top 5% Topper, 90% (Jan–Mar 2026).
Core — used across most projects
Python LangGraph FastAPI Groq LLaMA 3.3 Streamlit Git
AI / LLM
LangChain CrewAI MCP LoRA/PEFT PyTorch Hugging Face Transformers Whisper Prompt Engineering scikit-learn
Frontend & Data
React Next.js JavaScript/TypeScript SQL Supabase (pgvector) PostgreSQL Chroma pandas
Infra & Deployment
Docker Vercel Render Flask
- I built a multi-agent AI system that researches any startup in under 90 seconds
- I Built an AI Agent That Thinks Before It Answers — And Loops Back When It Doesn't Know Enough
- My LLM App Was Charging Rent-Controlled Tenants Penthouse Prices — So I Built a Router to Fix It
- Two Bugs That Almost Shipped in My Agentic RAG Assistant
- I Got Tired of My Portfolio Looking Like a List of Links. So I Built an MCP Server for It.
Have a problem worth an agent, a RAG pipeline, or an LLM integration? Reach out on LinkedIn, read the build logs on Dev.to, or start a contract on Upwork.