I build LLM products, then measure them.

AI and full-stack engineer in Mumbai, shipping five live products.

AI Expense Tracker99.7% of key fields right

The work, in 78 seconds.

Five products and the numbers behind them, in my own voice.

Five products, one rule: nothing a model says ships unchecked.

How each product moves from input to a checked result
Product	In	Model	Check	Out
AI Expense Tracker	Receipt or PDF	Gemini, strict schema	Decimal validators	Approved ledger
AI Phone Receptionist	A caller’s voice	Gemini, live audio	State machine	A booked slot
Deep Research Agent	A hard question	Parallel agents	Verbatim quotes	A cited report
DMI	Job listings	Local model draft	Human review	A tracked application
NotebookLLM Lite	Your documents	Local LLM, hybrid search	Citation check	An answer with its page

Projects, their measured results and links
Project	What it proved	Links
AI Expense Tracker	99.7% key-field accuracy; 88 of 88 planted errors caught	Live Code
AI Phone Receptionist	33 of 33 scenarios; 0 double bookings	Live Code
Deep Research Agent	21 of 22 answers correct; every citation quoted	Live Code
DMI	Explainable job matching; JWT auth and CI	Live Code
NotebookLLM Lite	Page-cited answers; no paid model APIs	Live Code

The agent can only book slots its search returned. Postgres blocks any overlap.

LLM features with an eval
Structured outputs, tool calls and agents, tested on a labelled set first.
Retrieval that cites
pgvector with full-text search; every answer points at its source.
Real-time voice
Streaming speech, barge-in and a state machine the model cannot break.
Schema to deployment
React, FastAPI, PostgreSQL and Docker, shipped with CI.