Dhruv Singhal — Bengaluru
I build things that think.
AI engineer. I spend my time on agents that don't wander, systems that survive contact with production, and the occasional business idea that has nothing to do with software. Mostly I like talking to people building interesting things.
IIT Kanpur · 3+ years building production AI · Open to collaborating
What I'm building.
Three things, walked through properly — the problem, the part that was hard, and how it behaves when it works.
AssistFlow
An agent that doesn't wander
Agents work in the demo and loop forever in production.
The happy path is easy. Real inputs are messy, the plan drifts, and nobody can tell you which step broke.
The scratchpad lives outside the model.
Long runs blow the context window, so working state is offloaded to an external store and pulled back on demand.
Step limits and loop detection.
When it starts going in circles, it gets cut off. An agent that stops cleanly is worth more than one that tries forever.
PromptShield
Attacking your own AI before users do
Most teams have never attacked their own AI.
Customer-facing LLM apps ship without anyone trying to break them. Then a stranger types the obvious thing.
Injection, extraction, guardrail bypass, role confusion.
A harness of adversarial prompts run against your system, reporting which categories it actually fails.
It runs on every commit.
Regressions get caught before users find them. A red run blocks the merge — that's the whole point.
What I'm building next
In progressA paper trading engine
thesis: mean reversion
why: 3 red closes, volume falling
thesis: mean reversion
why: 3 red closes, volume falling
An agent that forms a thesis and writes down why it was wrong.
It reads market data, takes a position on paper, and journals its reasoning so the decision can be re-read later.
Not whether it makes money.
Whether the reasoning survives an audit after the fact. That's a much harder and much more useful question.
No live capital.
This is a reasoning experiment wearing a trading costume. Nothing here is a strategy or advice.
If this is your kind of problem, I want to talk to you.
What I'm thinking about.
Why do most AI agents fall apart in week two?
The demo works. Then the inputs get weird, the loop wanders, and nobody can tell you which step broke.
Evaluation is the whole game, and almost nobody does it.
Everyone ships prompts. Very few ship a way to know whether today's version is worse than yesterday's.
The most interesting AI problems aren't in tech companies.
Trading desks, logistics yards, factory floors — still run on spreadsheets and phone calls, and the leverage there is enormous.
What happens when building software stops being the bottleneck?
If the code is nearly free, the scarce thing becomes taste, judgement, and knowing which problem is worth solving.
Poke around.
Type a command, or click one below. No AI, no tracking — just a terminal I wrote.
Good software is a series of decisionssomeone can still explain a year later.
Things I'd love to talk about.
If any of these are your thing, I'd like to hear from you.
Problems that aren't solved yet.
I've spent three years building LLM systems that businesses depend on — retrieval over private documents, agents that query real data, automation that runs unattended. IIT Kanpur, based in Bengaluru. I like problems that haven't been solved a hundred times already.