PROFILE
From complex questions
to deployed systems.
Applied AI engineer building LLM agents, MCP servers, and tool-driven workflows, from FastAPI backends to automated evals and cloud deployment. Focused on reliable agent execution: bounded tools, validated outputs, and deterministic fallbacks.
24Offline agent evals in CI
8.80/10B.Tech CGPA
2ndBRICS-FS-36 international final
02
Selected engineering work
TraceIntel
On-Chain Permission Risk Agent
Python · FastAPI · NVIDIA NOOA · OpenRouter · PostgreSQL · React · Docker · GitHub Actions
- Engineered a bounded LLM investigation agent with 3 read-only tools, a fixed call budget, and deterministic evidence validation; invalid output gets one repair attempt, then falls back to rule-based analysis.
- Built a pipeline that decodes ERC-20 approvals on Ethereum and Monad, comparing original allowances with current allowances, token balances, and spender contract code.
- Wrote 24 offline agent eval scenarios that run in CI, including 5 adversarial cases (prompt injection, out-of-scope tool calls, fabricated claims), separate from ~95 unit and browser tests.
- Implemented per-client rate limiting, daily LLM cost controls, and persistent agent execution traces for debugging.
Agon
Guardrailed Agentic Trading on Solana
TypeScript · MCP · Solana · Jupiter · Swig smart wallets
- Co-designed an agent-agnostic MCP trading server: AI clients research and paper-trade Solana tokens at live Jupiter prices through typed tools, with signing and execution kept outside the model.
- Designed agent-safety controls around a deterministic trade guard outside the LLM: per-trade spending caps, slippage validation, and signing only guard-cleared trades, never agent-written parameters.
- Defined fail-closed handling for uncertain trade outcomes: no blind retries, idempotency tracking, and explicit owner approval before trading permissions activate.
ARIS
Software Adoption Research Workflow
Python · Heym · NVIDIA NIM · Tavily (MCP) · Docker · DigitalOcean · Cloudflare
- Designed a hybrid AI workflow: 7 parallel research branches gather evidence (GitHub, OSV advisories, PyPI/npm, Stack Exchange, web search), LLMs extract structured findings, and the result is emailed as an adopt / trial / hold / avoid brief.
- Separated LLM research from deterministic decision-making, using Python to compute weighted scores, apply security vetoes, and calculate confidence scores.
- Built a Python execution node for the Heym workflow engine (7 tests, proposed upstream) and self-hosted it with Docker.
Study Buddy
MCP Server for Studying
Python · MCP SDK (FastMCP) · SQLite · Pydantic
- Built an MCP server for Claude Desktop with 13 tools, 3 resources, and 4 prompts over local notes, PDFs, and past papers, with SQLite mastery tracking and path-traversal protection.
More projects
03
Experience
SEP 2024 - NOV 2024 · KOLKATA, INDIA
Employability.life · RPA Developer & Business Analyst
Experiential learning program with Federation University Australia
- Automated manual Excel reporting with UiPath, saving 10+ hours per week and reducing errors by 70%; wrote process design documents and led requirements, testing, and sign-off across Agile sprints.
04
Technical skills
Languages & frameworks
Python · TypeScript · SQL · FastAPI · Pydantic · SQLAlchemy · React
Agents & LLMs
MCP / FastMCP · LangGraph · LangChain · LlamaIndex · OpenAI & Anthropic SDKs · NVIDIA NOOA · n8n
Models & retrieval
OpenRouter · NVIDIA NIM · Hugging Face · Ollama · pgvector · Chroma
Evaluation & testing
LangSmith · Langfuse · Ragas · Promptfoo · pytest · Playwright
Data & infrastructure
PostgreSQL · MongoDB · SQLite · Docker · GitHub Actions · AWS · Cloudflare · Vercel · DigitalOcean
05
Education
JUN 2022 - JUN 2026
Sister Nivedita University
B.Tech in Computer Science and Engineering
CGPA: 8.80 / 10
06
Recognition & credentials
DEC 2024
2nd Place · BRICS-FS-36
Data Analysis & Visualization. Awarded 2nd Prize in the international final held in China.
Certificate ↗FEB 2026
EL Excellence Award · Top 50 Learner
Federation University Australia