I build multi-agent AI systems that ask before they act.
Final-year CS undergrad and backend developer intern, spending most of my free cycles routing LLMs, wiring MCP servers, and putting a human in the loop before anything gets written.
I'm a final-year Computer Science undergrad at Vardhaman College of Engineering, Hyderabad, and currently a Backend Developer intern at Evernorth Health Services. Most nights I'm somewhere between a LangGraph state diagram and a Neo4j query console.
My focus has settled on agentic AI — systems where multiple specialized agents collaborate, but never take a real-world action (sending an email, writing a file, updating a calendar) without a human explicitly signing off. That pattern shows up across almost everything I build, and it's the closest thing I have to a design philosophy.
I also qualified GATE 2025 in Computer Science with an All-India Rank of 6450 (score 510) — proof that the theory behind the agents matters to me as much as the shipping does.
Working on backend systems within a large healthcare technology organization — building and maintaining the services that sit underneath production-facing products, with an emphasis on reliability and clean data flow.
Designing and shipping multi-agent systems end to end: routing logic, tool wiring, retrieval pipelines, and approval gates — treating each project as a small production system rather than a notebook experiment.
Coursework and self-driven study feeding directly into the project work below — GATE 2025 qualification (AIR 6450) along the way.
A LangGraph-based multi-agent workspace assistant with Gmail, Calendar, and Drive integrations wired through MCP servers — every write action stops at a human approval interrupt before it executes.
Five specialized CrewAI agents running sequentially over OpenRouter models to produce structured market research reports, with metrics tracking wired through the full crew execution.
A retrieval-augmented assistant answering university-specific questions, deployed and running live rather than sitting in a repo.
QLoRA continued pretraining of Qwen2.5-3B-Instruct into a clinical assistant, trained end to end on consumer-grade hardware.
A Neo4j + LangChain knowledge graph pipeline, batch-loading data via pandas and Cypher
UNWIND to work around Aura's hosted URL restrictions.
An SVM-based predictor wrapped in a Streamlit app — the project that started the habit of shipping every model as something clickable, not just a notebook.
Reaching out counts as one. Consider this the approval gate — send the request and I'll take it from here.