Product Manager · AI builder · Co-founder of an acquired startup

I find the real problem, then ship the product that solves it.

I'm a PM who owns the parts that decide whether a product works — the problem reframe, the strategy, the hard trade-offs — and builds the engineering when it's the fastest path to proof. I co-founded Vendaxo, acquired by Moglix, and have since shipped six products solo across very different problem spaces.

1
startup co-founded & acquired
6
products shipped solo
2
live in production
0→1
every one, from scratch
RAG & hybrid retrieval LLM orchestration Evals & safeguards AWS · production deploy B2B / marketplace Solo 0→1
Recognition
I co-founded Vendaxo, a marketplace for used industrial machinery — acquired by Moglix, the B2B commerce unicorn, in 2021.
The Economic Times ↗ Business Standard ↗ Entrackr ↗
Selected work

Six products, shipped solo

A body of work across AI, consumer, and tooling — one explored in depth below, the rest showing range and a steady shipping cadence.

AI · B2B marketplace Flagship

DTH / Mola

A production AI assistant that helps dental practices find and evaluate technology vendors — LLM orchestration, hybrid retrieval, an editorial data moat, evals and safeguards, deployed on AWS. The reframe: vendor selection is a trust problem, not a search problem.
484-vendor moat · 2–3s response · a founder raised capital on it
AI · Consumer PWA

Sayless

A live PWA of AI-scored speaking-practice drills. The meta-story: a PM who spotted his own weak spot — spontaneous communication — and shipped an app to train it.
11 milestones · 208 practice tiles · 70+ tests
AI · Career tooling

CareerAssistant

An AI-assisted system for running a focused job search — targeting the right roles and landing the next one.
AI · LLM-powered
Collaboration tooling

Viscollab

A real-time collaborative document workspace with AI-assisted editing and review.
Real-time · AI-assisted
Developer utility

MarkDownCopy

A Chromium browser extension that copies any web page's content as clean Markdown.
Browser extension · shipped
Simulation

InventorySim

A simulation project — another shipped build demonstrating a consistent cadence of getting things out the door.
Shipping cadence

CareerAssistant, Viscollab, MarkDownCopy, and InventorySim aren't the headline acts — they're the evidence of range and cadence. I don't just plan products; I ship them, repeatedly, across different problem spaces.

A closer look · flagship · GRIP

DTH / Mola, in depth

How the flagship was reasoned about and built — the decisions and trade-offs, not the code.

Live · search.dentaltechhub.com
The reframe everything turned on

Vendor selection isn't a search problem. It's a trust problem. Practices weren't short on options — they were drowning in 480+ of them with no one credible to match them. That single reframe set the product, the data strategy, and the safeguards.

G — Gap

The problem

Dental practices were drowning in 480+ vendor options with no trusted matcher. Every "directory" was a list; none answered the real question — which vendor is right for a practice like mine, and can I trust the answer?

  • Choice overload, zero credible guidance
  • High-stakes, expensive, sticky purchases
  • Trust gap, not an information gap
R — Result

What shipped

A live AI assistant in production with a zero-hallucination eval baseline — credible enough that the founder raised capital on the platform.

  • Live at search.dentaltechhub.com
  • Zero-hallucination eval baseline
  • Investor-credible — raised on it
I — Insight

What I learned

In a trust product, your safeguards are your feature. Accuracy isn't a backend concern you bolt on — it's the thing the user is actually buying. So the eval suite and safeguards got first-class roadmap status, not "later."

  • Trust must be engineered, then measured
  • An eval baseline is a product asset
  • The data moat is the durable advantage
P — Plan

How it was built

A deliberate sequence: build the data moat first, choose an LLM-conductor architecture, make safeguards a feature, then engineer latency and cost to production-grade — kept current by a refresh pipeline.

  • Data moat → architecture → safeguards
  • Then latency & cost engineering
  • Then keep the corpus fresh
The engineering depth — decisions & trade-offs (for the technical reader)+
Latency journey — time-to-first-content (TTFC)
38s
v1 baseline
15s
after streaming
2–3s
production now
Held to $0.03–0.05 per turn — a real unit-economics budget, not a demo
1

Build the editorial data moat first

Curate 484 vendors as structured, editorial-grade data. In a trust product the corpus is the product — no clever model rescues a thin dataset.

Trade-off: Slow manual editorial work up front instead of a fast scrape. Chose durability and credibility over speed-to-demo — the moat competitors can't quickly copy.
2

Replace the state machine with an LLM conductor

The first design was a hand-built state machine — brittle, every new path meant new branches. I rearchitected to an LLM conductor that orchestrates retrieval and tools dynamically.

Trade-off: Gave up the determinism of explicit states for flexibility and far less maintenance — and bought control back with tight safeguards and evals.
3

Make accuracy retrievable: hybrid retrieval

Pair semantic search (Chroma) with keyword search (BM25) and fuse rankings with RRF. Semantic alone misses exact product names; keyword alone misses intent.

Trade-off: More moving parts vs. a single vector store. Worth it — grounded retrieval is what makes the zero-hallucination baseline achievable.
4

Treat safeguards as a feature, not a filter

A 7-module safeguards pipeline backed by 79 tests and an eval suite. "It didn't make something up" is the headline feature — so it gets tested like one.

Trade-off: Engineering time with no visible feature. But it's the difference between a demo and something a founder can raise money on.
5

Engineer latency and cost to production-grade

Drove TTFC from 38s → 15s → 2–3s (streaming + retrieval/orchestration tuning) while holding $0.03–0.05 per turn.

Trade-off: Constant tension between quality, speed, and token cost. Set an explicit per-turn budget and optimized within it.
6

Keep the moat alive: corpus-refresh pipeline

A repeatable refresh pipeline (on AWS Elastic Beanstalk + RDS) keeps the 484-vendor corpus current. A stale data moat stops being a moat.

Trade-off: Ongoing operational cost vs. a one-time dataset. Treated freshness as a product commitment — trust decays with every outdated fact.
LLM-conductor orchestration
Hybrid retrieval · Chroma + BM25 + RRF
484-vendor editorial moat
7-module safeguards · 79 tests
Eval suite · zero-hallucination baseline
AWS Elastic Beanstalk + RDS
Writing · How I think out loud

Notes on product, AI, and building

Essays on product, communication, money, and building — here and on my Substack, Pages to Practice.

Contact

Get in touch

Always happy to talk product, AI, or building things that ship.

Email
nirat.adi@gmail.com
LinkedIn
linkedin.com/in/niratp ↗