I'm a second-year CSE student at IIIT Sri City. I build agent systems and the tooling around them: MCP servers, RAG pipelines, static analysis, serverless backends. Most of it is solo, built in a 24-hour sprint or a weekend, against real APIs and real rate limits.

The part I care about most is where the model is not in the loop. Wherever an answer has to be correct, I use code that can be tested, and I bring in an LLM only for what code can't do. I lean on AI heavily while building. I decide the architecture and I check the output.

Demo video · Live app (email sign-in) · Code · Technical writeup

Under India's MSMED Act, a small supplier is owed compound interest at three times the RBI Bank Rate on late payments. Almost nobody claims it, because computing it correctly is fiddly. DueClaim takes the invoice, works out the exact amount owed with a month-by-month statement, flags the buyer's income-tax exposure under s.43B(h), and generates the demand notice as a PDF.

• There is no LLM in the product. The interest engine is pure Python on Decimal. Interest is never stored; every read recomputes it, so the dashboard and the notice can't disagree.
• It refuses to guess. An invoice whose interest would start before the earliest RBI rate on file is rejected with a clear error instead of being priced wrongly.
• OCR only proposes. Textract extracts invoice fields with confidence scores, and nothing is written until the user confirms it. Manual entry always works.
• 194 backend tests, 25 of them on the engine. Fully serverless on AWS, built solo in one weekend for the WeMakeDevs × AWS First Commit hackathon, deployed for under $2.

Python FastAPI Lambda API Gateway DynamoDB Cognito Textract S3 SAM React TypeScript

flowchart LR
    B["Browser<br/>React · Vite · TS"] -- "JWT" --> G["API Gateway<br/>Cognito authorizer"]
    B -- "presigned PUT" --> S["S3<br/>invoice uploads"]
    G --> L["Lambda<br/>FastAPI + Mangum"]
    S --> L
    L --> E[["Interest engine<br/>pure Python · Decimal · no I/O"]]
    L --> D[("DynamoDB<br/>single table")]
    L --> T["Textract<br/>AnalyzeExpense"]
    style E fill:#0a3d62,stroke:#00d9ff,color:#fff

A local MCP server and CLI that turns a repository into a code graph (Tree-sitter, NetworkX and local Git history), so an assistant can see callers, callees and blast radius before it touches anything.

• Seven MCP tools: analyze_repo, search_graph, trace_flow, find_issues, get_metrics, visualize, get_git_context.
• Evidence over confidence. When a call can't be resolved it says unresolved or ambiguous instead of inventing an edge, and a scoped scan is labelled partial.
• It never executes the repo it analyzes. Parsing, graphing and Git inspection all run locally.
• 4,900-line engine, 1,570-line test suite, CI on a Python 3.11/3.12 matrix with ruff, a wheel build and an end-to-end MCP smoke test. Packaged as a plugin for Claude Code, Codex and Antigravity.

MCP Tree-sitter NetworkX Git pytest GitHub Actions

Classifies each message as notify, digest or mute. Built for HackerRank Orchestrate, where it placed #376 of 1,983.

• Deterministic detectors for scams, phishing, prompt injection and domain impersonation run before any model call. They can't be prompt-injected, and they settled 24 of 110 messages without one.
• TF-IDF retrieval with reaction boosting over 400+ historical messages, personal context for 54 users and 23 groups, Whisper for 13 voice notes (some in Hindi), a vision model for 19 of 20 images, and Llama 3.3 70B for the final call.
• Requests are paced at 2.1 s to stay under Groq's 30 RPM limit, which gave zero rate-limit errors on free tiers.

Llama 3.3 Whisper v3 Nemotron VL TF-IDF Groq OpenRouter

Small changes, but each one removed a real stumbling block for the next contributor.

agents      LangGraph · MCP · multi-agent orchestration · tool-calling
llm / rag   Groq · Gemini · Anthropic · Whisper · ChromaDB · MiniLM · TF-IDF
analysis    Tree-sitter · AST · NetworkX · Git history
backend     Python · FastAPI · PostgreSQL · REST
cloud       AWS Lambda · API Gateway · DynamoDB · Cognito · S3 · Textract · SAM
ml          PyTorch · ResNet · SAR/optical fusion
frontend    React · TypeScript · Vite · Tailwind
tooling     pytest · GitHub Actions · Docker · ruff · Linux
languages   Python · C/C++ · JavaScript/TypeScript

• HackerRank Earn: invited to the expert network as a top performer at Orchestrate.
• Shaastra 2026 (IIT Madras): 4th in Quantified Dilemma, finalist in 6 other technical competitions.
• TEDx IIIT Sri City: Domain Lead, Speaker and Program Curator. Also in the Entrepreneurship Cell (operations) and the IOTA Club (AI engineering domain).

I'm looking for remote AI/agent engineering roles and internships, research collaborations, and open-source work. If you're building something agentic that has to work outside a demo, email me.