Sohil Pathan · full-stack & AI engineer. From the React screen people click to the multi-agent backend, with the evals and tracing in between.
class Sohil:
role = "Full-Stack & AI Engineer"
builds = ["multi-agent systems", "MCP tool layers", "RAG", "LLM evals", "React + FastAPI apps"]
learning = ["agent failure analysis", "AI-assisted dev tooling", "whatever ships next"]
believes = "the model is an unreliable component, so engineer around it"
night_shift = True- Research assistant for financial advisors. For an enterprise client, I built a supervisor–worker agent system on AWS Bedrock with MCP tool servers, metadata-filtered RAG (80% lower retrieval cost) and Langfuse/Comet evals, delivered as a FastAPI + React app. Research that took 5–8 hours now takes about 10 minutes.
- Legacy code migration with agents. I designed six LangGraph agents with AST transforms and human approval gates to move EJB code to Spring Boot. In the proof of concept, 95% of migrated code compiled and passed tests.
- Safe LLM automation for servers. I built a Python SSH framework where the model can only choose from an approved command list. Pre-upgrade checks dropped from 3–8 hours to about 1 hour per server, across 400+ servers.
- Agents & AI: LangGraph, LangChain, Semantic Kernel, MCP, RAG, AWS Bedrock, SageMaker
- Evals & tracing: Langfuse, Comet
- Backend: Python, FastAPI, Node.js, Express
- Frontend: React, Next.js, TypeScript, Tailwind CSS
- Data & cloud: PostgreSQL, MongoDB, ChromaDB, AWS, Azure, Docker
- 🏆 Won Thomson Reuters' global AI hackathon (100+ teams)
- 🎓 Claude Certified Architect – Professional (CCA-P)
- 🥈 2nd place, Azadi Ka Amrit Mahotsav Hackathon 2022 (2000+ teams in Gujarat)
- 💡 SSIP 2.0 government grant for an NFT marketplace for cultural artifacts
⚡ Fun fact: I keep night shift switched on at all times.