answerd.it builds AI for the parts of the economy that still run on spreadsheets, phone calls, and thirty years of a person's judgment. We do not make horizontal chatbots. We ship one deep tool at a time, aimed at one trade at a time. Currently: construction estimating.
Describe a construction project in plain English. TakeoffAI returns a full line-item takeoff with regional unit costs, a bid strategy with three scenarios and their win probabilities, and a graded history of every bid it has ever written for you. It is a local install. Bid data never leaves the building.
Describe the project in plain English — "9,200 sq ft retail TI, demo the old build-out, new framing, electrical rough-in, HVAC revisions". TakeoffAI parses the scope, pulls regional unit costs, applies your overhead and margin, and returns a full line-item takeoff. No templates, no RSMeans subscription.
Once the estimate is on the table, BidToWin reads the RFP and hands back three scenarios with win probabilities attached, a list of scope gaps the owner is going to ask about on the walk-through, and a range of where competitors are likely to come in. You still make the call. You just make it with the rest of the board visible.
Every estimate runs as a tournament: five estimator
personalities (conservative, balanced,
aggressive, historical_match,
market_beater) across three temperatures, sampled
twice each. Thirty calls in parallel. The grid collapses to one
consensus bid per personality — the one closest to the
median — and those five are what you see. Disagreement
between them is itself a signal.
When a job closes, TakeoffAI records what each personality predicted against what the job actually cost, and tracks a rolling Brier score for its win-probability calls. Any personality that drifts past 5% average deviation over its last five closed jobs gets red-flagged and sidelined from your next tournament. Accuracy is a measurement, not a claim.
Upload your old bids — won and lost. Every closed job
becomes a page in a local Obsidian vault, and the
historical_match personality pulls comparable jobs
from your own history into the prompt for the next bid in the
same trade. After the fiftieth bid it sounds less like a
generic estimator and more like the person who trained it.
Still experimental, still the most interesting piece: if one
personality starts winning more than 60% of your tournaments,
the harness evolver takes the trace logs, reasons about why the
others are losing, and rewrites their prompts in place. A
dry-run mode shows you the diff before anything ships. Rollback
is git revert.
| Scenario | Bid | Win prob |
|---|---|---|
| Conservative margin | $358,200 | 41% |
| Balanced — recommended | $344,938 | 67% |
| Aggressive | $329,400 | 82% |
| Competitor range | $312k – $371k | n=4 |
Historical match flags three comparable retail TIs you won between $332k–$351k, all with 11–13% markup. Aggressive gets the job 82% of the time but loses money on a third of those runs based on your last twelve closed jobs. Balanced is the one your company has actually won with.
Every closed job is a graded prediction. The ones that drift get benched.
A short answer to the question every contractor's IT person asks before a tool gets on their machine.
backend/data/client_profiles/<client_id>.json.
Every contractor grows their own.
wiki/jobs/, automatically
enriched and linked.
Currently in field testing with GCs in Central Texas.
