← All posts
product 8 min read

From a paragraph to a live scanner in ten minutes

A walkthrough of the model builder — describe a setup in plain English, watch it compile, replay-validate against 3000 bars, activate.

Every trader has a setup that isn't on the shelf. It's the pattern you noticed in your journal three months ago, or the confluence you keep sketching on paper, or the specific mix of two ICT concepts that works for you but nobody else seems to teach.

Sentio's model builder exists so you can turn that pattern into a live scanner without writing code. Here's what a full session looks like, from opening the builder to a model firing alerts.

1. Describe plain English 2. Clarify AI Q&A 3. Compile DSL spec 4. Replay 3000 bars 5. Refine chat + re-run 6. Activate live scanner iterate until stats meet your bar
Six steps, one loop. Activation is gated on the replay pass.

Step 1: Describe the setup

Open /model-builder. There's a text field and a chart upload button. You describe the setup in whatever detail you can:

When ES sweeps the London high on 15m and NQ fails to confirm, wait for a bearish displacement candle and enter on a 5m FVG retracement during the New York session. Stop above the swept high, target 2R.

That's a real setup. The builder handles it. If you attach a chart screenshot the vision model reads it too — useful for setups that are easier to point at than to write.

Step 2: Sentio AI asks clarifying questions

Before compiling, the AI reads your paragraph and asks the questions it needs to disambiguate. For the setup above, it might come back with:

  • "Sweep the London high" — do you mean the exact high of the London session (00:00–07:30 UTC), or the highest swing during that window?
  • "Fails to confirm" — does that mean NQ made a lower high, or that NQ didn't sweep its own equivalent liquidity?
  • "Bearish displacement" — minimum size relative to average range, or just any red candle?

You answer in plain English. The AI restates its understanding. You confirm, adjust, or clarify. Two or three turns and you're done.

Step 3: The spec compiles

Once the AI has enough, it compiles your description into a DSL spec — a structured list of primitives that Sentio's engine knows how to run. Something like:

  • smt_divergence → bearish, ES → NQ, 15m
  • liquidity_sweep → london_high, 15m, ES only
  • displacement → down, 5m, min_impulse ≥ 1.2× ATR
  • fvg → retest, 5m
  • entry → session ny, execution 5m
No hallucinated primitives

Every primitive is a real building block in the engine, not a fabrication. If the AI wants to compile your description into something the engine doesn't support, it stops and asks you to rephrase. There's no "fake it and hope no one checks."

Step 4: Replay validation

Here's the gate that separates a hypothesis from a live model. Before you can activate the draft, it has to run against replay data — ~3000 historical bars on the pair you're targeting.

The validator plays the model through the tape and reports:

  • How many times it triggered
  • The R distribution (mean, median, best, worst)
  • The win rate at T1 and T2
  • The worst intraday drawdown
  • Sample screenshots of the highest-confidence triggers
Hard activation gate

If the model triggers fewer than 5 times in ~3000 bars, it's too thin to trust and Sentio blocks activation. If it triggers 50+ times with negative expectancy, you're looking at your model and deciding whether to refine or abandon. Either way, you know what you're activating before you activate it.

Step 5: Refine with follow-up chat

Not happy with the replay results? Open the model's chat and describe what you'd change. "The triggers are too frequent — require a stronger displacement." The AI compiles the change into an updated spec, you re-run replay, you compare.

The chat is per-model and persists across sessions. Six months from now you can come back to a model and see the exact conversation that produced its current spec. That's the audit trail signal groups don't give you.

Step 6: Activate into the live scanner

Once you're satisfied with the replay validation, one button activates the model. It joins the scanner alongside the built-in models, produces the same shape of setups (entry, stop, targets, reasoning trail), and gets subject to every discipline rule you've written.

That last part matters. Custom models aren't second-class. If your daily cap trips at trade 3, it applies to your custom model too. If your revenge-cooldown timer is running, the custom model waits like every other. The whole system enforces the same rules regardless of which model surfaced the setup.

What this replaces

The traditional path from "I noticed a pattern" to "I have a scanner for it" is: learn Pine Script, buy a TradingView subscription, spend a weekend fighting v5 syntax, deploy it, discover it fires on every green candle because your definition of "sweep" was ambiguous, spend another weekend fixing edge cases. Ship, watch it fail live, iterate.

Sentio replaces that with: describe, clarify, replay-validate, activate. Ten minutes end-to-end for a first draft. And the primitives library is real code that's been used by hundreds of scans, so you don't have to reinvent "what counts as displacement" — you inherit the same definition every other model uses.

See the model builder page or sign up free and build one yourself.