About
Final-year CS engineer shipping production AI with evaluation wired in from the start. At Deloitte, took a LangGraph agent from in-request execution to a self-hosted agent server with durable PostgreSQL state, proven live on SAP at 8 of 8 tasks. Across 32 public repositories, most of them deployed and free to use, every reported number is reproducible from the repository, and three of them ship as installable packages on PyPI.
Experience
-
AI Engineer (Agentic AI) InternDeloitte India, Digital Excellence Centre · Gurugram, IndiaJun 2026 – Aug 2026
Migrated SAP Agentic, Deloitte's system that automates SAP configuration, from LangGraph running inside a FastAPI request handler onto a self-hosted Aegra agent server, with threads, runs and checkpoints in PostgreSQL inside the firewall at no licence cost. Set the migration's acceptance test as field-by-field identical output between the original and migrated paths, and held the public API unchanged so no consumer needed modification. Proved the new path on the live SAP tenant: a Bank Accounting goal was planned into 8 configuration tasks, executed in a real browser through Playwright MCP, recovered from a duplicate-entry conflict and a mis-filled field, and finished 8 of 8 in about 15 minutes. Traced two defects that returned empty results with no exception, a state field overridden by a LangGraph reserved name and a streaming path discarding every chunk, to their root cause in the installed packages' own data models.
Education
-
B.Tech, Computer Science and EngineeringThe LNM Institute of Information Technology, Jaipur · LNMIIT2023 – 2027
-
Class XII (Higher Secondary), CBSEDon Bosco · CBSE2021 – 2023
-
Class X (Secondary), ICSEH.P. Children's Academy · CISCE (ICSE)2019 – 2021
Skills
Tools / apps / platforms
Languages
Projects
-
Python, SQL, Entity Resolution, DBSCAN, Zoho Catalyst
Built end to end over 50 days for the Karnataka State Police. Joined 59,985 FIRs across 31 districts and 298 stations into one link graph with entity resolution, a 0.870 AUC three-month forecast, DBSCAN hotspots, evidence OCR and a grounded English and Kannada assistant. Ranked 26,168 flagged cases into one worklist with a reason and next action for each.
-
LightGBM, CNN-LSTM, DuckDB, Google Cloud Run
Forecasts air quality with LightGBM on station data plus a CNN-LSTM over a 15,360-cell satellite grid, ranks interventions by modelled return, and checks whether each dispatched order worked using difference-in-differences against control regions. Served from one embedded DuckDB file on Cloud Run.
-
LangGraph, MCP, Playwright, Amazon Bedrock, Aegra
Built the same planner five times (linear, parallel, orchestrator, human-in-the-loop, MCP factory graph) and kept every generation runnable. Served all five concurrently through Aegra on AWS Bedrock with a standalone MCP server exposing 10 travel tools, and 630 backend tests against a real PostgreSQL.
-
Node.js, MCP, LLM Agents, Zod, Vitest
One platform where LLM agents turn a URL and plain-English intent into API tests and security checks, so what an agent does against a live endpoint can be accounted for.
-
Python, FastAPI, Embeddings, Redis
A drop-in proxy that keeps the OpenAI request and response shape, including streaming, verified against the official Python and Node SDKs and LangChain, and answers reworded repeat questions from a semantic cache. Published on PyPI.
-
Python, Amazon Bedrock, Agent Loop
Ranked #12 in a 24-hour agentic challenge. The agent proposes payment plans and a deterministic layer decides: a 90-day cash-flow forecast keeps only plans that never breach the minimum balance, and the ranking is re-applied after every model call.
🎯 Hobbies & interests
- Debating
- Model United Nations
- Open-source building
- Cricket
- Badminton
- Kabaddi