ANKIT KUMAR · FOUNDING ENGINEER

Making complex
things feel
simple.

I build real-time voice AI and agent systems, and the production infrastructure that keeps them fast, safe and affordable. Web, mobile and backend.

Explore my work Say hello

Currently building at Optima Learn /Noida, India

THERE’S MORE DOWN THE ROAD

A LITTLE CONTEXTSmall team.
Production stakes.

10K+students learning on the platform

2.5smedian spoken reply from the AI coachSpeech → reasoning → speech, live in production.

60s+ → <20sworst-case AI coach cold start

01 / SELECTED WORK

Systems that hold up
in production.

A few things I’ve taken from a question
to something people can use.

OPTIMA LEARNIN PRODUCTION

SpeechReasoningResponse

OPTO · A PERSONAL AI COACH FOR EVERY ASPIRANT

01 · PRODUCT ENGINEERING · 2025 - NOW

Optima Learn

Personal AI coach
Voice AI
LLM observability
Production reliability
Mobile

Opto, a personal AI coach.
For every student.

An AI-native CAT exam-prep platform used by 10K+ students. I lead its engineering: the architecture, Opto (a personal AI coach for every student), the production infrastructure, and the iOS/Android app. Most of the work is the part after the demo: latency, cost, safety, and staying up when traffic does something unexpected.

Next.js
NestJS
PostgreSQL
Prisma
WebSockets
OpenAI
Sarvam STT/TTS
Langfuse
GCP
Terraform
React Native

Visit the platform

Inside the build Decisions, constraints, and outcomes

A PERSONAL COACH

A coach that knows the student.

Designed and built Opto end to end: agent architecture, student context, long-term memory, consent-driven actions, safety and evals. The goal is coaching that stays consistent across weeks of preparation. Shipped the experience across web and a React Native app for iOS and Android, including sign-in, in-app purchases and secure token storage.

Tool calling
Context engineering
Long-term memory
Safety & evals
React Native

HANDS-FREE VOICE

A conversation that can keep up.

Opto listens and talks back in real time: streaming speech-to-text, a reasoning model and sentence-level text-to-speech over NestJS WebSockets, with barge-in so students can interrupt, and auto-reconnect. Holding one speech stream per session made our vendor’s stream limit the real ceiling, so I moved to per-utterance leases with a Postgres-backed limiter shared across autoscaled Cloud Run instances. That removed a hard 20-concurrent-user limit. Median spoken reply is 2.5 seconds.

Voice AI
WebSockets
Streaming STT/TTS
Concurrency control
Postgres

PRODUCTION SYSTEMS

Reliable, observable production.

A timed race and immediate status feedback cut worst-case cold-start replies from over 60 seconds to under 20. Traced a crawler surge to database connection-pool limits and rewrote founder-dashboard reporting in SQL, cutting ~190 queries to ~50 per load. End-to-end Langfuse tracing and a per-call cost ledger make voice turns, speech calls and prompts visible, so we can track quality and spend per feature and per student.

Incident debugging
Connection pooling
SQL performance
Langfuse
Cost tracking

ASKCODE

any / github-repository

01Understand the files
02Embed the summaries
03Retrieve relevant context
04Stream a grounded answer

RETRIEVAL FLOW / ILLUSTRATIVE DIAGRAM

02 · DEVELOPER TOOLING · JAN 2025

Askcode

RAG
Retrieval design
Grounded answers

A conversation with
your codebase.

Ask a GitHub repository questions in plain English. Instead of embedding raw code, Askcode has an LLM summarise each file and embeds the summaries, so search matches what the code does rather than what it is named. Answers stream token by token and decline anything the retrieved files cannot support.

Next.js
tRPC
PostgreSQL + pgvector
LangChain
Gemini
Vercel AI SDK

Open Askcode

Why summaries, not source? A closer look at the retrieval decision

THE APPROACH

Search by meaning, not by name.

LangChain loaders walk the repository and an LLM summarizes each file. Those summaries become 768-dimensional embeddings in PostgreSQL, so similarity search works with a description of the code’s purpose.

THE GUARDRAIL

Stay inside the evidence.

Cosine-similarity search assembles top-k file context for token-streamed answers through the Vercel AI SDK. Prompt-level grounding asks the model to decline questions the retrieved context cannot support.

02 / THE JOURNEY SO FAR

Not a straight line.
A good road, though.

A few stops that shaped how I build.
Education, experiments, and the work between.

FIVE CHAPTERS. ONE SCENIC ROUTE.

✦01
012021 - 2025
Learning the fundamentals.
Maharaja Agrasen Institute of Technology
B.Tech in Information Technology
B.Tech in Information Technology at MAIT, Delhi, graduated with 80%. The degree gave me the theory; the side projects taught me how to ship.
- Information Technology
- C++
- Delhi
✦02
02Sep 2023 - May 2024
First users, first data problems.
GEMS of JCI
Software Development, Research & Data Analysis Intern
Built features for a social platform in a team of three: React/Redux on the front, Node.js and MongoDB behind it. Rewrote slow reports with aggregation pipelines and indexes, cutting data-processing time by 20%.
- React
- Redux
- Node.js
- Express
- MongoDB
✦03
03Jan 2025
Learning retrieval by building it.
Askcode
Personal project · RAG for codebases
What if exploring a codebase started with a conversation? A RAG system that searches file summaries instead of raw code and answers only from what it retrieves. My first real AI project.
- RAG
- pgvector
- LangChain
- Gemini
✦04
04Jun - Aug 2025
Shipping to real students.
Zarle Infotech
Software Engineer · Agency work on Optima Learn
Worked on Optima Learn as an agency engineer, then joined full-time as its Founding Engineer. Built Next.js/TypeScript interfaces with accessibility in mind, wired Firebase Auth, Firestore and Cloud Functions, and shipped production releases on Vercel.
- Next.js
- TypeScript
- Firebase
- Vercel
✦05
05Sep 2025 - PresentNOW
Leading the build.
Optima Learn
Founding Engineer
I lead engineering on an AI-native exam-prep platform with 10K+ students. I set the architecture across web, API and mobile, built Opto, a personal AI coach for every student, and run the production side: incidents, security and observability.
- Engineering leadership
- Voice AI
- AI agents
- Production systems
- GCP

The chapters overlap, as good journeys tend to. Education ran alongside the early work.

03 / A BIT ABOUT ME

Curious by default.
Engineer by doing.

Ankit Kumar in the mountains — Ankit Kumar

Hey, I’m Ankit - kankit on the internet.

I’m a software engineer based in Noida, working across product interfaces, backend systems, and applied AI. I like understanding how the pieces fit together, then taking responsibility for the whole thing.

At Optima Learn, that means moving between a student-facing interface, a voice pipeline, a database migration, and a production issue. The tools change; the job is to make the experience work.

My background is in Information Technology at MAIT, Delhi. Since then, I’ve been building across the stack, with a growing focus on the reliability and behavior of AI-powered products.

The full résumé

THE TOOLKIT