The Model Was Never the Product. The Harness Is.

Ask a trader what they want and not one of them will say “a large language model.”

They'll say: show me the coins coiling for a breakout. Show me stocks being quietly accumulated into the close. Show me what had the most volatile weekend. The distance between that sentence and a correct answer is the entire business. Closing that distance has a name — a harness — and I've become convinced that building harnesses is the actual end state of all applied AI.

That's what Sortefi is: a natural-language market screener. You type what you mean, in plain English, and it hands you the symbols — across crypto, US equities, Indian equities, ETFs, and forex. No Pine Script. No 40-field filter form. No “learn our query language first.” Just intent in, watchlist out.

The problem I actually solved

Screening has always been a tax on the trader. The pros pay it in Bloomberg terminals and custom code. Everyone else pays it in rigid dropdown screeners that can express “RSI below 30” but fall apart the moment you want something a real strategy sounds like:

“Three consecutive green 2-hour bars where the average body is wider than the last fifty, during the weekend.”

Try building that in a dropdown. You can't. You'd write code — if you can code. So the market split into two camps: pros who could express sophisticated ideas and retail who couldn't. I didn't want a smarter chatbot. I wanted to collapse that gap — to give a retail trader the expressive power of a quant desk through the one interface every human already has: language.

That's the whole thesis of Sortefi as a natural-language stock and crypto screener. Say the strategy. Get the names. The sophistication moves into the engine, not onto the user.

Why the model is the smallest part

Here's the uncomfortable truth people building on top of AI eventually run into: the model is a commodity, and it's the easy 15%.

When you screen on Sortefi, the language model does two things — it turns your English into a structured intent, then it writes the Python that expresses your filter. That's it. That's the glamorous part everyone talks about. Then the harness takes over, and the harness is the company:

• A market-data engine — millions of candles across five markets, minute-resolution, cleaned, time-aligned, and correct across every timeframe.
• A library of 120+ indicator functions the generated code is allowed to call — the real API surface of the product.
• A sandbox that executes model-written code safely against live data.
• Routing, warmup math, and completeness guarantees so the answer isn't just plausible — it's right, and it doesn't silently drop half the universe.
• A verification loop that catches the model when it's wrong and makes it try again.

Delete the two AI calls and hardcode the Python, and ~85% of this system is still there — and still the hard part. The intelligence is necessary. But intelligence on its own is inert. Value shows up the moment you wrap it in a system that grounds it in real data and verifies its output. That wrapper is the harness.

The harness is the end goal of all AI

Look at where every genuinely useful AI product is heading and you'll see the same shape: not a smarter conversationalist, but a tighter loop between intent and verified action against a real system. Coding agents aren't valuable because they can talk about code — they're valuable because they run it, read the error, and fix it. The best AI tools don't answer; they do, then check that what they did actually worked.

A raw model is a brilliant intern with no hands, no memory of your data, and no way to know if it's wrong. A harness gives it hands, grounds it in truth, and closes the loop. As models commoditize — and they will — the durable value migrates entirely into the harness: the data, the tools, the guardrails, the verification. The moat was never the model. The moat is everything the model's output runs on.

Anyone can wire up an AI call over a weekend. Almost nobody can stand up clean minute-level data across five markets with correct multi-timeframe math and completeness guarantees underneath it. That's the part that took the work. That's the part that lasts.

What comes next

The same engine that answers “most volatile weekend” can answer questions retail traders have never been able to ask a screener before — the ones pros pay dearly for:

• Volatility term structure — where short-horizon volatility has spiked above its own baseline (something is happening now).
• Trend quality, not just direction — the clean, holdable trends versus the choppy drift to the same place.
• Idiosyncratic movers — names moving on their own catalyst, not just riding the market.
• Conditional-probability edges — “after an N-sigma down day, how often is the next three days green?” A real, quantified statistical edge, expressed in a single English sentence.

Every one of those is a screen a hedge-fund analyst wants and a retail trader has never had a way to run. The harness makes them all a sentence away.

The model gave us language. The harness turns language into an answer you can trade on.