Boopathi Raja

I ship. 4 systems in production. AI agents, self-play agent for miniBR, POS for real shops. Numbers are true.

[Boopathi Raja @Boopathiraja]

@HeydariAI @kaggle In production, scaling compute is necessary but not sufficient: the bottleneck is usually eval coverage, orchestration, and reliable deployment. The systems that ship treat compute as an input and build the feedback loops that turn it into consistent agent behavior.

[Boopathi Raja @Boopathiraja]

@karpathy This points to evals moving from static prompts to task-completion traces. In production we pair screenshots and trajectories with invariants—correct stock, tax, and receipt—because an agent can produce a beautiful artifact while still failing the business state.

[Boopathi Raja @Boopathiraja]

This matches what we see in shops: the highest-leverage AI engineer is the one who watches the cashier’s workflow and changes the spec before writing code. We ship when the agent handles messy inventory and payment edge cases with evals, not when it merely generates a plausible PR.

[Boopathi Raja @Boopathiraja]

@swyx The interesting production shift is that the agent owns the loop, not just code generation. In our shipped systems, it earns trust only when every tool call is observable, retries are bounded, and an eval catches the wrong write before it reaches a shop’s ledger.

[Boopathi Raja @Boopathiraja]

@demishassabis @ShaneLegg The interdisciplinary angle matters in production too: our POS agents need product, finance, and ops constraints in one eval—not just a language benchmark. Shipping gets safer when each workflow has an explicit owner, fallback, and audit trail.

[Boopathi Raja @Boopathiraja]

@sama Worth waiting if the last mile is right. The production test we use for POS agents is end-to-end: inventory lookup, tax, payment state, receipt, and retry path all pass evals—not just a polished demo.

[Boopathi Raja @Boopathiraja]

@OpenAI A useful production extension is to log the tool call, permission path, and final side effect—not only the model trace. In our POS/agent flows, evals pass only when the receipt is correct and the fallback is safe, even when the model sounds confident.

[Boopathi Raja @Boopathiraja]

I don’t tweet vibes. I ship systems shops run on.

4 in production: steel trading billing, clothing retail, bakery POS + kitchen tickets, govt exam prep.

Also building Nexq (queue/booking): 76 Supabase migrations, 41 edge functions, 31 Playwright specs.

Portfolio: https://t.co/0dzjA8Bqoy
GitHub: https://t.co/wfSgDKDAND

[Boopathi Raja @Boopathiraja]

Built a Safe-to-Spend workbook for feast-or-famine freelancers: trailing "salary," tax vault (educational, not advice), buffer runway, one safe-to-spend number. Founding $19. Feedback welcome.
https://t.co/3yGg6zHv2X

[Boopathi Raja @Boopathiraja]

I'm smoke-testing a tiny Freelancer Client Ops Kit: preloaded Notion-style client pipeline + proposal/invoice templates for solo freelancers who hate overbuilt CRMs. Founding $29. Roast the positioning if it's weak.
https://t.co/7Xenmq6dm4

[Grok @grok]

Grok Build now gets better the more you use it.

It remembers conventions, decisions, and project facts across sessions.

/memory to browse
/dream to organize recent notes into topics

https://t.co/A2MMZT5eJX

[Boopathi Raja @Boopathiraja]

Better grok @bot is cost effect and more usecase i think 🤠

[maria @maria_rcks]

$4K FOR A SUB??? and they said muse was cheap

[maria @maria_rcks]

$4K FOR A SUB??? and they said muse was cheap

[Boopathi Raja @Boopathiraja]

Paper fleet recap (simulation only). Built and run with Grok Bot.

Parent + 50 workers on Polymarket-style crypto Up/Down. Combined sim equity climbed into seven figures. Win rate ~47%. Real prices, fake fills.

Live money is a different story. My Bangalore livebot burned real USDC on short crypto bets and is paused near empty. Paper ≠ live.

Not financial advice. Research / build log, powered by @bot

#Polymarket #PaperTrading #Quant #GrokBot

[Boopathi Raja @Boopathiraja]

@fuyu_kasumii yo

[Boopathi Raja @Boopathiraja]

Every time @thsottiaux announces reset it moves weekly actual reset far behind and feels like he actually gives reset (4) that what we suppose to use in a month

[Boopathi Raja @Boopathiraja]

@melvindvivas Feels like @thsottiaux should give reset every 3 days once untill they understand to give real 20x usage on Max

[Boopathi Raja @Boopathiraja]

@boltdotnew Forge