Skip to content

AI Builder & Engineer

Chinmoy Paul

Chinmoy Paul

I build AI systems that hold up in production

An AI builder who takes a product from a rough idea to something deployed, and keeps it running.

Last role
AI Engineer Intern
Nistula · May–Aug 2026
Studying
B.S. Data Science & AI
IIT Guwahati · 2027
Based in
Bangalore, India
Open to remote
Last shipped
Acumen
the backtest that said no

Where I Have Worked

Experience

AI Engineer Intern

Nistula · Boutique villa hospitality · Goa, India

May – Aug 2026

Three months owning the AI layer end to end — the guest-facing assistant, and the booking site behind it.

What I shipped

WhatsApp AI host

Nistula Assistance

Built solo to a tagged v1.0.0: a seven-tool agentic Claude loop that quotes live prices, reads PMS bookings, dispatches staff jobs against per-kind SLAs and remembers each guest across stays. Seven deterministic guardrails then run after the model and before every send — so the assistant structurally cannot state a price no tool returned, or promise an action nobody performed.

1,813 tests · 17 blocker defects caught pre-merge · 62-item live PMS backlog drained, 0 errors

  • TypeScript
  • Fastify
  • Postgres
  • pg-boss
  • Claude Sonnet 4.5
  • WhatsApp Cloud API

Direct-booking villa site

Nistula Website

Replaced a brochure page that could not take a booking. It reads live availability and pricing from the eZee PMS behind one swappable adapter, takes Razorpay payment and writes the reservation back — with the server re-quoting every price so a value from the browser is never trusted, and idempotency keys on all writes.

15 PMS endpoints behind one adapter · a Zod schema each · every price re-quoted server-side

  • Next.js 15
  • React 19
  • Drizzle
  • Neon Postgres
  • Razorpay

My Portfolio

My latest Work

Agentic systems, evaluation harnesses and the infrastructure around them. Each one was built against a written spec and sealed by an independent review.

Acumen — ten-year NSE backtester

Data Engineering · Quant

Acumen

A backtester and live screener that replayed one trader's intraday NSE F&O strategy over ten years of minute data and proved, with a full audit trail, that it loses money. The whole repo exists to make an unwanted answer hard to disbelieve.

188,345 trades · 2,621 tests, 0 failures · 28 adversarial reviews

  • Python
  • pandas
  • Parquet
  • pytest
Samadhan — AI civic resolution agent

Agentic AI · Live

Samadhan

An AI civic agent that takes a complaint from a single photo to verified resolution — it files in the citizen's own language, tracks the statutory SLA, and autonomously drafts the escalation ladder up to an RTI when the deadline lapses.

Live on Cloud Run · 36 tests, green CI · Open311 export

  • Next.js
  • Genkit
  • Gemini 2.5 Flash
  • Firebase
Nistula Assistance — WhatsApp AI host

Agentic AI · Production

Nistula Assistance

A WhatsApp AI host for a Goa villa company — live-priced quotes, staff job dispatch with SLAs, and an automated guest lifecycle. The model writes; seven deterministic checks decide whether what it wrote may leave the building.

1,813 tests · 7 code-enforced guardrails · 0 model-invented prices

  • TypeScript
  • Fastify
  • Postgres
  • Claude Sonnet 4.5
LLM on Kubernetes, instrumented for reliability

Infrastructure · Observability

LLM on Kubernetes

An open-source LLM on a single-node cluster where the interesting problem isn't getting it to run — it's knowing continuously whether it's still producing tokens fast enough. Ollama exposes no metrics, so I built the exporter.

8 metric families · 11 Grafana panels · alerts that fired on real degradation

  • Kubernetes
  • Prometheus
  • Grafana
  • Ollama
Smart Returns Resolver — bilingual support triage

LLM Orchestration · Live

Smart Returns Resolver

Bilingual EN/AR support triage where an LLM reads the complaint and a hand-written 36-cell policy table decides the entitlement. The classifier returns its own escalation verdict — and the engine deliberately ignores it.

88.5% on a 78-dimension eval rubric · 0 pipeline failures

  • Next.js
  • TypeScript
  • Zod
  • gpt-4o-mini
Identity Reconciliation Engine

Backend · API

Identity Reconciliation Engine

One HTTP endpoint that collapses a customer's scattered emails and phone numbers into a single identity. Instead of walking a graph, it holds the link depth at exactly one and makes merges pay for it — union-find with eager path compression.

2 SELECTs per lookup, no matter how many merges

  • TypeScript
  • Express 5
  • Prisma
  • PostgreSQL

Also built

Nistula — direct-booking villa website

Nistula Website

A direct-booking villa site wired live to the eZee PMS behind one swappable adapter, with Razorpay payment and server-side re-quoting on every write.

  • Next.js 15
  • Drizzle
  • Razorpay
Live site
Thine Qualifier — scored signup funnel

Thine Qualifier

A seven-question funnel that scores every signup and ranks the operator's outreach list. Two questions score backwards on purpose — it measures unmet pain, not seniority.

  • React
  • Vite
  • Airtable
GEO Tracker — LLM brand visibility

GEO Tracker

Asks Gemini the questions a real buyer would type, then scores whether a brand survives into the answers — a surface conventional SEO tooling cannot see.

  • Python
  • Streamlit
  • Gemini
Repository
Cone Trading — the atomic order path through a Spring Boot trading backend

Cone Trading

A crypto trading backend: users sign up, fund a wallet through Razorpay or Stripe, buy and sell at live CoinGecko prices, track a portfolio and withdraw to a bank. 37 endpoints across 14 entities, with JWT auth and email two-factor.

  • Java 17
  • Spring Boot
  • MySQL
Repository
Newsletter Brief Generator — three article URLs rendered into one email-safe HTML brief

Newsletter Brief Generator

Three article URLs in, one email-client-safe HTML brief out. Two of its four scrapers hang off generated CSS hashes rather than stable hooks, and the README says which ones will break first.

  • Python
  • BeautifulSoup
  • Jinja2
Repository
Show More

What I Have Written

Writing & Code

I write up the systems I build — including the parts that did not work.

Featured · Hashnode · Apr 2026

How to Deploy an Open-Source LLM Reliably

3,400 words on why a healthy pod tells you nothing about a working model — and the two-layer readiness split that fixed it.

Read

Code · GitHub

Every repository behind this page

Each project with its spec, decision log and review trail committed alongside the code.

View profile

Connect with me

Get in touch

Happy to walk through any of the repositories, including the parts that did not work.