Adarsh Prashar

Backend/infra engineer · Codeforces CM (top 100 IN). Building RiskKernel — open-source kill switch + budgets for AI agents.

[Adarsh Prashar @prashar32]

If managing or maintaining the portal is difficult, there are many capable technology partners who can help. As a software engineer, I'm even willing to help identify and fix the issues for free if that's what it takes to make this service reliable for everyone.

Please fix this urgently. Citizens depend on it during emergencies.

[Adarsh Prashar @prashar32]

@police_haryana @gurgaonpolice @cmohry @Cyberdost
@DoT_India
Please try using your own complaint portal before asking citizens to rely on it.

My phone was stolen, and I'm simply trying to file a complaint so I can obtain the complaint number required to submit a CEIR phone-blocking request.

I've been trying for over an hour, but the website repeatedly breaks in the middle of the complaint process. This is a critical public service, and it shouldn't fail when people need it most.

https://t.co/LfAByoVbqz

[Adarsh Prashar @prashar32]

Building agent reliability infra as a solo engineer. OSS in Go by day, live trading system at night.

Next deep-dive - which one do you want?
a) the risk engine internals (how code overrules the model)
b) crash-resume design (how agents survive dying mid-run)

[Adarsh Prashar @prashar32]

Every agent framework this year is converging on the same three primitives: state, tools, and a human-in-the-loop hook.

Which tells you the framework was never the hard part.

Orchestration is a solved problem now. What happens when the orchestrated thing does something you didn't want - still mostly unsolved, still mostly bolted on after the fact instead of designed in from the start.

[Adarsh Prashar @prashar32]

Debugging a distributed system and debugging a wrong greedy proof are the same muscle.

You have a belief. Reality disagrees.
Find the smallest case where your belief breaks.

[Adarsh Prashar @prashar32]

The meta-lesson after 2 months of live trading:

The only reliable way to run a non-deterministic system is inside a deterministic one.

I've been building that principle into a reusable open-source layer — happy to share details if you're wrestling with the same problem, just ask.

[Adarsh Prashar @prashar32]

I built a multi-agent AI trading system solo in ~2 months.

It's been live on a real broker with real money for 2 months.

Here's the architecture, and the 3 decisions that kept it from blowing up:

[Adarsh Prashar @prashar32]

Bonus: two models debate every trade.

Disagreement is signal. When they split, size down or skip.

An ensemble is the cheapest confidence estimate you'll ever get.

[Adarsh Prashar @prashar32]

Decade of competitive programming, ~4000 problems, CM on Codeforces (top 100 India).

LLMs made none of it obsolete. They made it the whole game.

Code is cheap now. Knowing exactly what to build, in what order, and why it's correct — that's the entire job that's left.

[Adarsh Prashar @prashar32]

80–90% of my code is AI-written now.

What changed isn't speed. It's that reviewing became the job.

Writing code was never the bottleneck. Trusting it is.

[Adarsh Prashar @prashar32]

Question for people running agents in production:

what actually stops your agent when it goes wrong?
a) a budget enforced in code
b) a human approval gate
c) vibes

Honest answers only. Mine was (c) for an embarrassingly long time.

[Adarsh Prashar @prashar32]

A network timeout and a broker rejection look identical to your agent.

One means "retry."
The other means "you're about to double your position."

If your agent touches money and your orders aren't idempotent with client-side IDs, you're not running a system. You're running an experiment.

[Adarsh Prashar @prashar32]

I ran a 6-agent trading system on real money for 2 months.

The agents were often wrong. The system never was.

The difference: the risk engine is compiled code the agents can't override.

LLMs for judgment. Code for limits.

[Adarsh Prashar @prashar32]

-> Every AI agent demo: "look what it can do".

-> Every AI agent in production: "look what it can't do anymore".

The second one is the product.

[Adarsh Prashar @prashar32]

Repo:
https://t.co/NVzA37z5eZ

I’d love feedback from people running agents in real systems:

What would you need before trusting an agent runtime around production workflows?

[Adarsh Prashar @prashar32]

I’ve been building RiskKernel as the SRE layer I wanted before putting AI agents near production.

Agent frameworks help agents reason.
RiskKernel makes the run bounded, resumable, auditable, and governable.

Open-source, self-hosted, Apache-2.0:
https://t.co/NVzA37z5eZ

[Adarsh Prashar @prashar32]

The goal is not to replace LangChain, CrewAI, AutoGen, Langfuse, LiteLLM, or observability tools.

It sits between them.

Frameworks orchestrate reasoning.
Observability shows what happened.
RiskKernel enforces the run boundary.