M Lakshmanan

30+ yrs enterprise software. Building my own trading systems with AI. Design, not tips.

[M Lakshmanan @muthulakshman]

Every "System One model" article uses the same demo: a support ticket routed to billing vs technical vs sales.

Here's one doing a real job instead. A decision-only model (Jev) and an LLM running together in production, every morning, in a system I actually depend on 🧵

[M Lakshmanan @muthulakshman]

One day of real data, so a progress note, not a result.

They differed on at least one of four fields in 21 of 27 items. High enough that I'm watching, not celebrating. Sonnet 5.5 running alongside as a third opinion.

Anyone else splitting work this way?

#Jev #AIEngineering

[M Lakshmanan @muthulakshman]

Every "System One model" article uses the same demo: a support ticket routed to billing vs technical vs sales.

Here's one doing a real job instead. A decision-only model (Jev) and an LLM running together in production, every morning, in a system I actually depend on 🧵

[M Lakshmanan @muthulakshman]

So they split the work. Haiku reads and writes. Jev files and says how sure it is. No winner to pick.

Not a fair fight on classification, to be clear: Jev chooses from a fixed list, Haiku was producing JSON from a prompt.

[M Lakshmanan @muthulakshman]

None of these fixes were clever. Each is something I'd have insisted on in a team a decade ago.

Building solo with AI, you move fast enough to skip the discipline you already have. Nothing pushes back until money is involved.

What's the green light you trust most?
#AlgoTrading

[M Lakshmanan @muthulakshman]

I run a personal automated trading system on Indian markets. Alerts from TradingView, a Python server making the decisions, orders to my broker.

Its worst failures have never been crashes. They're green lights that lie: process running, port answering, nothing actually working 🧵

[M Lakshmanan @muthulakshman]

The watcher that kept fixing my tunnel was half the problem. Two processes repairing the same thing is how 65 minutes happen.

Now one process repairs and a separate watcher only watches, asking one question every few minutes: did an alert TradingView fired reach my database?

[M Lakshmanan @muthulakshman]

@realSamHu Been there. For an hour our tunnel sent every alert to a dead port while the monitor kept "fixing" it. Now preflight checks effects too. A real API call per token, check the response body. And a watcher pings me on Telegram if a fired alert never reached the server

[M Lakshmanan @muthulakshman]

@NiravExplains That’s pretty good.Good luck

[M Lakshmanan @muthulakshman]

Not new in software. It's plain indirection. Just rare in webhook trading setups.

Next: what happens inside the server between the alert and the order.

How do you structure your alerts?

#AlgoTrading #TradingView

[M Lakshmanan @muthulakshman]

Real crash? A different config sells everything at market. Same alert, one word changed.

Need a different split? The alert overrides one number.

Defaults live in config. Exceptions travel with the alert.