Aankit Roy
Book a conversationBook

AI strategy and production systems

AI strategy from someone who has built it at scale.

I advise founders, CEOs and boards on where AI creates real value, then help their teams build it. Twelve years of engineering leadership, from a Y Combinator startup I co-founded to platforms serving millions.

  • 5M+users on Khabri
  • YC S19co-founder & CTO
  • 3M+paying users at Writesonic
Portrait of Aankit Roy
Aankit RoyCo-founder, stealth AI startup

Where I’ve built

  • GeoIQ
  • Isha Foundation
  • Writesonic
  • iMumz by Pruoo Healthcare
  • Khabri
  • Y Combinator

Results

Measured in production, not in a demo.

Three numbers from systems I led, each one taken from real traffic: latency people felt, cost the business stopped paying, and scale that held.

Writesonic↓ 73% latency

1,500ms

Response time for 3M+ paying users, with infrastructure cost halved.

Before1,500 ms
After400 ms

GeoIQ↓ 75% cost

−0%

Agent cost, on agentic systems handling 10,000 sessions a day in production.

Before100%
After25%

Khabri, YC S1914 ms responses

0.0M+

Listeners on a vernacular audio platform, served with 14 ms response times.

014 ms100 ms

Track record

Built for millions, before advising anyone.

2026 – now

Stealth AI startup

Co-founder

Building an autonomous AI engine.

Nowbuilding in stealth

2025 – 2026

GeoIQ

VP, AI Engineering

Built the AI engineering team from scratch and shipped agentic systems in production, while cutting agent cost by 75%. Presented the work at the Ai4 conference.

10kagent sessions a day, in production

2023

Writesonic

Engineering Manager

Took response times from 1.5s to 400ms for 3M+ paying users and halved infrastructure cost.

3×the engineering team, tripled

2024

Pruoo Healthcare (iMumz)

Head of Engineering

Rebuilt the engineering team behind a million-user parenting platform.

−80%app launch time

2025

Isha Foundation

Engineering Leader

Led engineering on the Miracle of Mind project.

2017 – 2023

Khabri, Y Combinator S19

Co-founder & CTO

Built a vernacular audio platform to 5M+ users with 14ms response times.

$3.2Mraised, including from Y Combinator

About

Engineer first, advisor second. The advice I give comes from systems I have shipped and teams I have led, not from slides.

I started out automating lending workflows for banks, co-founded Khabri through Y Combinator, managed engineering at Writesonic and led AI engineering at GeoIQ. Today I help leadership teams choose the right AI bets, then stay close enough to the build that the strategy survives contact with production.

The full story
12
years building software people use daily
4
million-user platforms built or led
20
engineers led as CTO of Khabri
Ai4
conference talk on agents in production

Current thinking

Decisions you can put a threshold on.

Most of what an AI agent does all day is small decisions: route this, flag that, escalate or don’t. I design systems where every decision comes back with an honest probability, so the business chooses what runs on its own and what a person sees.

Read the essay on Jev
One shared context is encoded once, three typed questions read from it in parallel, and each returns a probability for every allowed option.Shared context, encoded oncekeys and values cachedWhich queue?payments.81risk.15ops.04Is it urgent?yes.64no.36Needs a person?yes.22no.78Three answers, each a probability over its allowed options
Fig. 1 Decide, don’t talk. The context is read once; typed questions branch from it in parallel and return calibrated probabilities instead of paragraphs. Five questions over a 2,000-token context need roughly 4.7× less prefill than five separate calls.
Reliability diagram. A calibrated model tracks the diagonal, so a 0.8 answer is right about 80% of the time. An overconfident model sits far below it.a person reviewsacts on its own0.00.00.20.20.40.40.60.60.80.81.01.0Stated confidenceHow often it was rightCalibrated modelConfidence written as text
Fig. 2 A threshold only works if the probability is honest. Calibrated answers above 0.8 run on their own; everything below goes to a person. Illustrative curves.

In conversation · ETCIO

On leading technology for millions of listeners.

A video interview with ETCIO on the role of a technology leader, and what it took to scale Khabri’s platform.

  1. The role of a technology leader
  2. Scaling Khabri’s platform to millions

Press play to watch

Testimonials

Aankit’s technical architecture decisions at Writesonic helped us scale from handling thousands to millions of AI requests daily. His optimization work reduced our infrastructure costs by 73% while improving performance dramatically.
Samanyou GargFounder & CEO, Writesonic
“We scaled to 6 million users and achieved $1.2 million ARR over 6+ years… He consistently delivered tech solutions that exceeded industry benchmarks, even under the typical constraints of an early-stage startup.”
Pulkit SharmaCo-founder of Khabri, YC alum
“His tech skills are unparalleled, and he is an innovative thinker. His binary approach to decision-making is spot-on.”
Dushyant KohliCo-founder & CEO, Khabri
“His leadership style is characterised by a unique ability to motivate and keep the enthusiasm of the entire team high.”
Pulkita MarwahGoogle Cloud Consultant, Searce

For founders, CEOs and boards

If AI is on your board’s agenda, let’s talk before the roadmap is set.

  1. Where AI can create value in your business
  2. What it would take to build it

Fig. 0 · LiveA confidence surface, drawn as you watch. The bright line is the 0.80 threshold: above it a decision runs on its own, below it a person decides.