OPEN
SPSumit Parmar
Ahmedabad, IN--:--
AI Engineer · Ahmedabad, India

Sumit Parmar

From what AI can do, to what it actually does.

Most AI stays a promise. I build the systems that turn it into real work, from the first requirement to production.

Product · Full stack · LLM systems · Agents · Infrastructure
Sumit Parmar
SCROLL
01 · About me

Hi, I'm Sumit.

AI engineer at Pucho.ai, and in the final year of an Integrated Master's in AI and ML. I work directly with clients, turn what they need into product designs, and build the systems behind them, from the first prompt to the server it runs on. I also guide the junior engineers on our team.

P/01

Smart is easy. Reliable is engineering.

Getting a model to answer is one prompt. Getting it right every time, in production, is the real work.

P/02

Every token is a cost decision.

The right model per task, on quality, latency and price, with routing kept explicit.

P/03

AI breaks silently.

Wrong routing, wrong prompt roles, wrong date logic. You only catch it if you look.

02 · What I do

One engineer, the whole path.

From the first client conversation to the server it runs on. These are the layers I own.

03 · Featured projectIn progress

Enterprise Brain

A company brain that connects mail, tasks and code, and answers like someone who was in every conversation. Then agents act on it.

HOW IT WORKS · ANIMATED
Key decisions
HydraDB as the base

Evaluated against pgvector for this workload.

A graph only where it pays

Owners, decisions and dependencies get structure. The rest stays in vector memory.

Time and permissions travel with the data

Answers respect who can see what, and when it was true.

PythonFastAPIHydraDBKnowledge graphMCP
Also at work

Autonomous dev loop. Agents that ship, not just suggest.

A task comes in. The agent plans, writes the code, runs the tests, fixes what broke, opens the PR and takes it all the way to GitHub. Tests must pass, and human approval is a setting per repo.

ONE TASK, END TO END · ANIMATED
04 · Selected workSwipe

More systems in production

Built for live clients. Open any card for the full trace.

05 · Projects

Things I build on my own.

Open source work with real code, real numbers and a demo you can run yourself.

Open source · Agent evalsLive

evalloop

Tests for LLM agents that run every case more than once.

Pass rate100%was 33%
Reliability100%was 0%
p95 latency1.0swas 3.1s

Demo support agent, simulated with planted bugs, 5 trials per case.

PythonZero dependenciespass@k and pass^kGitHub ActionsGroq, OpenAI, Ollama
06 · ResearchBuilding

Reinforcement learning for LLMs

Fine-tuning makes a model sound right. RL with verifiable rewards makes it be right. I build the environments where that can be proven.

ONE TRAINING STEP · ANIMATED
SIMULATED TO EXPLAIN THE METHOD · NOT RESULTS
01SampleThe policy writes a group of answers to one task.
02VerifyHidden checks score each answer.
03CompareEach answer is judged against its own group.
04UpdateBetter answers become more likely.
Go deeper: the math, and what makes an environment trustworthy
A · Group relative advantage
Ai = ( ri − mean(r1..G) ) / std(r1..G)

No separate value model. An answer is rewarded for beating its siblings on the same prompt, which keeps training stable and cheap.

B · Objective, kept close to the start
max E[ Ai · clip(ρi) ] − β · KL( πθ ‖ πref )

The KL term stops the policy drifting too far from the reference model while it chases reward.

C · What makes an environment trustworthy

Expert-written tasks

From people who know the domain, not scraped from text the model has already seen.

Hidden verifiers

Programmatic checks the model never sees. Reward comes from what is true.

A sealed benchmark

A held-out set that is never trained on. If the score moves there, it is real.

Reward hacking checks

Every environment is attacked on purpose before it is trusted.

GRPOPPORLVRReward designVerifiersPyTorch
07 · Experience

Where I've shipped

NOW

Currently

08 · Contact
Let's ship something real.
sumit@portfolio: ~ · interactive

Open to conversations about production AI, retrieval and agent systems. WhatsApp is the fastest way to reach me.

UP/DOWN navigateENTER openESC close