ai eng & final year student. i build agentic llm systems and the full-stack products around them, because an agent nobody can use is just a research script.

now

callzie, an ai voice agent that phones customers, negotiates a new appointment time, and commits the rebooking through live tool calls before the call ends. next.js 16, typescript, postgres, retell ai, on google cloud run.

built

• callzie, an ai voice agent that calls customers about their appointment and rebooks it mid-conversation, writing the new slot to the database while they are still on the line. double-booking is impossible at the database layer, via a postgres exclusion constraint rather than check-then-write app logic that two calls can both slip through at once.
• eval-loop, a self-improving agent: generate, evaluate, critique, revise, then measure whether it actually got better. the judge is validated against human labels first, so improvement is real and not the judge drifting.
• discourse-mcp, an mcp server from scratch that lets any llm app search & synthesize what developers are actually saying across hacker news, lobsters & reddit. 11 typed tools written to the protocol directly, so the same server drops into claude desktop, an ide, or a langgraph agent unchanged, and answers come back cross-source rather than from whatever one forum thought that week.
• dermwise, skin-lesion classifier (efficientnet-b0, 75% val macro f1) with a qlora fine-tuned tinyllama writing the report. macro f1 because ham10000 is imbalanced and plain accuracy would hide the rare classes.

stack

langgraph · langchain · mcp · rag · llm eval · pytorch · qlora · faiss · typescript · next.js · react · node · python · fastapi · postgres · google cloud · docker

away from the screen

i love to cook, sketch and paint. art is everything and everything is art.

portfolio · linkedin · anushapundirwork@gmail.com

open to ai/ml & full-stack internships.