"I've shipped enough broken AI projects to know — the architecture meeting you skip always becomes the bug you can't find."
I'm an Agentic AI Developer from Gujarat, India. I build autonomous AI systems — RAG pipelines, multi-agent graphs, LLM-backed CLIs — and the one thing every project I've shipped has in common is that I planned the structure before writing a single line of code.
That habit didn't come from a course. It came from building things without it and feeling exactly what breaks and why. Now every project starts with a clear execution layer, defined state boundaries, and deliberate tradeoff decisions — before a file even gets created.
I work across the full AI engineering stack: from protocol-level MCP tool servers and LangGraph stateful graphs, to FastAPI backends and Streamlit interfaces that real users can actually interact with.
Currently building Capsule — a personal content manager with a Telegram bot frontend, FastAPI backend, Groq-powered processing, and hybrid SQLite + ChromaDB storage.
📍 Open to AI Engineer roles at AI-first startups — Mumbai, Pune, Bangalore, Hyderabad, or remote.
🔬 Multi-Agent Research Pipeline | 🔗 Repository
Autonomous CLI system executing structured, stateful research workflows across 6 specialized AI agents.
- Hierarchical Orchestration: Designed a full supervisor-driven agent hierarchy — Supervisor → Search → Scrape → Summarize → Critique → Synthesize — with explicit state boundaries between each role to prevent context bleed across agents.
- Quota-Aware Model Routing: Built dynamic fallback logic using Groq (llama-3.3-70b) as primary and Gemini (gemini-2.0-flash) as fallback tier — agents switch models mid-pipeline without interrupting the workflow.
- Human-in-the-Loop Checkpoints: Engineered interactive review gates at critical pipeline stages with a clean Rich terminal UI, keeping a human in control without breaking the execution flow.
- Stack: Python, Groq, Gemini, DuckDuckGo, Rich CLI
🧠 hArI — RAG Document Intelligence | 🔗 Repository
Production-deployed web app for grounded, hallucination-resistant document interrogation.
- Precision Retrieval: Switched ChromaDB distance metric from L2 to cosine and enforced a strict
SCORE_THRESHOLD=0.35— completely eliminated hallucinated source citations without touching the LLM layer. - Intent-Based Routing: Built a local intent classifier that decides query path before hitting the vector store — Conversational → Vector RAG → LLM fallback — reducing unnecessary embedding lookups.
- Clean Stream Output: Wrote a custom
strip_thinking()post-processor to scrub raw LLM reasoning tokens before they reach the UI, keeping responses clean without modifying the model behavior. - Stack: Groq (llama-4-scout-17b), ChromaDB, SentenceTransformer (all-MiniLM-L6-v2), PyMuPDF, Streamlit, uv
🤖 AI Agent Engine | 🔗 Repository
A 4-layer autonomous agent pipeline built natively in Python — architected before a single file was created.
- Zero-LLM Routing Layer: Designed a deterministic cache at the top of the pipeline that resolves ~80% of routine queries with 0 LLM API calls — speed and cost handled at the architecture level, not prompt level.
- Sub-50ms Semantic Search: Integrated ChromaDB + SentenceTransformer maintaining ~30ms semantic search latency as the second routing layer before any external API call is made.
- Strict Execution Economics: Planner → Validator → Executor pipeline with hard quota enforcement keeps per-session cost at ~$0.0005 — a constraint that was designed in, not optimized in later.
- Stack: Python 3.11+, Gemini API, ChromaDB, SentenceTransformer, DuckDuckGo, Open-Meteo
| Project | What it does | Stack |
|---|---|---|
| 🔌 DevMind — MCP Server | Local Model Context Protocol tool server giving LLMs secure, HITL-gated access to the file system — read, write, execute Python snippets, format JSON, count tokens | Python, MCP SDK, tiktoken |
| 🗄️ LangGraph SQL Runner | Multi-question parallel SQL execution using LangGraph's Send API for dynamic fan-out across schema analysis and execution nodes, with inline HITL review before any query fires | LangGraph, Groq, SQLite, Pydantic |
| 💰 Finance Agent CLI | Terminal-based personal finance assistant — 8 natural language commands, zero cloud retention, all transaction data stays local | Python, Groq, JSON, CLI |
| 🎯 NextSteps | Resume-to-JD gap analyzer — parses unstructured resumes against job descriptions or URLs and outputs skill mapping + actionable roadmap | Python, Groq, Tavily |
Capsule — A personal saved content manager built the right way: schema first, then logic, then interface.
- Frontend: Telegram bot + browser extension for saving content from anywhere
- Backend: FastAPI with async endpoints and Pydantic-validated request/response models
- Processing: Groq for content summarization and tagging at save-time
- Storage: SQLite for structured metadata + ChromaDB for semantic search across saved content
Building this because every content manager I tried either had no AI or had AI bolted on. This one is designed around it.
I'm actively looking for AI Engineer roles at startups where AI is the product, not a feature. If that's you — or you know someone building that — reach out.
