The Pattern We Keep Building: Turning Scattered Signals Into One Clear Recommendation
Standup text, home-walkthrough video, and zoning ordinances look nothing alike. Here's the one AI pattern Vereda's studio keeps reusing across all three products — and the one place we deliberately don't use AI at all.
Ask what Vereda AI, PreLoss, and ADU Home Resources have in common, and the honest answer isn't "AI" — that word is doing too much work to mean anything on its own. The real answer is narrower and more useful: all three take something a person already has, that's too messy or too scattered to act on directly, and turn it into one thing they can actually use.
An engineer's standup update is a paragraph of freeform text. A homeowner's walkthrough of their living room is two minutes of shaky video. A zoning question is "can I build this here," and the honest answer lives in a PDF on a county website nobody wants to read. None of that is usable as-is. That's the actual product in all three cases — not "AI," but the compression from scattered input to one clear output.
We didn't set out with a thesis and go looking for three industries to apply it to. We built Vereda AI first, noticed the pattern while building the other two, and it's worth naming so it's not an accident going forward.
The pattern, named
Three ingredients, every time:
1. The input is already there, and it's a mess. Nobody has to be talked into writing a standup update, recording a walkthrough video, or asking whether they can build an ADU. The signal exists. It's just unstructured — free text, video frames, a question with no obvious source of truth.
2. A model reads the mess and extracts structure. Not "AI magic" — a specific, narrow extraction task. What does this standup response imply about blockers and sentiment? What items appear in these video frames, and what are they worth? What does this specific city's ordinance actually say?
3. The output is one clear thing, not a pile of maybes. A burnout risk flag, not a wall of sentiment scores. A line-item inventory, not a folder of screenshots. A sourced answer, not five search results to sort through yourself.
The hard part is never step 2 — GPT-4o vision and GPT-4o-mini are both good enough at extraction that "can the model do this" stopped being the interesting question a while ago. The hard part is steps 1 and 3: picking an input people were already producing anyway, and being disciplined about what "one clear thing" actually means for that specific person's decision.
Three examples, one shape
Vereda AI: the input is a standup response someone typed in under two minutes. The extraction is sentiment, blocker mentions, and engagement patterns against that person's own baseline. The output is a single flag — this person's trend is worth a conversation — instead of a manager re-reading eight people's updates every morning trying to spot the pattern themselves. We wrote about how that detection actually works if you want the deeper version.
PreLoss: the input is a video walkthrough of a room, shot on a phone. The extraction is GPT-4o vision reading frames pulled from that video. The output is a line-item inventory with replacement values and a PDF report — not "here are some photos of your stuff," but a structured document an insurance adjuster can actually use. We'll go deeper on the pipeline itself in an upcoming post.
ADU Home Resources: the input is a homeowner's question — can I build an ADU on this lot, in this city. The extraction is a model that searches for that city's actual zoning ordinance instead of answering from memory. The output is a specific, sourced answer with real numbers, not "check with your local planning department." We'll walk through exactly how that works next week too.
Different industries, different models, same shape: take what someone already has, and hand back the one thing they were actually trying to figure out.
Where we don't use AI on purpose
The pattern also tells you where *not* to reach for a model, and that restraint matters as much as the pattern itself. ADU Home Resources has an ROI calculator — projected rental income, cap rate, break-even timeline, cash flow against a loan. That's just arithmetic. A mortgage formula doesn't get better because a language model produces it instead of a function; it just gets a chance to be subtly wrong in a way that's harder to check. So the calculator is plain deterministic math, no model involved, because deterministic math is more trustworthy than a model guessing at a formula it could just compute directly.
That's the actual discipline: use extraction where the input is genuinely unstructured and a model is the only way to get structure out of it. Don't use it as a layer of polish on something that was already a calculation.
Why this is a studio thing, not a product thing
None of this required three separate companies. It required noticing that "read something messy, hand back one clear answer" is a pattern worth having opinions about — where AI genuinely earns its place, where it's overkill, how to build the extraction step so it fails safely instead of confidently making something up.
That's also, practically, what our AI consulting work is: the same judgment calls we make on our own three products, applied to whatever messy input your team is already sitting on. If you're trying to figure out whether a problem in your own product is a "build an extraction pipeline" problem or a "just write the function" problem, that's a conversation we're happy to have.