Use casesResearch at scale
Research no team could do by hand, with a source on every figure.
Some lists are worth a fortune and nobody has one: every school that trains the people you hire, every supplier in a category, every site that matches a rule. We build research agents that find what you missed, read every page, record a source for every figure and rank the results by what matters to you. No model grades its own work.
What gets in the way
Hand & Stone runs 651 spas, and every one of them runs on licensed massage therapists. The most direct source is the schools that train them. Its list of those schools was built by hand, thin on detail and probably missing schools.
They wanted it longer and deeper: eleven researched fields on every school and a score that says where to start. That is research at a scale no team does by hand.
And a confident wrong answer sends a recruiter to the wrong door. A model that writes fluent nonsense is worse than no list at all.
What we build
Agents search and read. Plain code does everything that has to be exact.
- Discovery agents work region by region, each shown what is already known so it hunts for what is new. Public datasets add candidates vetted by someone else, and code checks every find against your list by fuzzy name.
- A screening agent asks one question in two searches, does this place do the thing at all, and rules out the rest before any deep research is spent.
- Research agents investigate each survivor, several at once, and return a strict structured record with a source for every figure. Finished work saves at once, so a crash loses nothing.
- A priority score in ordinary code, with the rubric written out and every weight yours to change. Missing data is never scored as zero. Below a known threshold, a record goes unranked instead of guessed.
- Your original spreadsheet comes back intact, with the new columns and the new records added beside it.
How we know it works
We check the agents without asking a model.
- A separate step fetches every page each record cites and looks for the name and the figure itself. For Hand & Stone, the right school was named on 1,230 of 1,239 checkable records, and 71% of the figures appeared on the page.
- Where research disagrees with a public filing, the filing wins. Rereading the page behind each screening verdict reinstated three schools the agent had wrongly ruled out.
- The weak spots are written down. Tuition was quoted per credit, per hour or per program, so the workbook says to open the source before quoting it.
- The finished database held 1,193 schools, 572 of them new, and 68 of the 201 top-tier schools were not on the original list. Every one of the 2,424 agent runs was metered by stage.
Client work
1,193 schools in the finished database
572 new schools found beyond Hand & Stone’s original list
68 of 201 top-tier schools that were not on the original list
Read the story
Where it fits
Franchising
AI for franchise brands, researched at a scale no team does by hand.
Questions
How do you know the agents did not make things up?
Every figure carries a source, and a separate step with no model in it fetches that page and checks the name and the figure are on it. Records that fail are flagged, not quietly kept.
What does it cost to run?
It is metered by stage, so you see what discovery, screening and research each cost. Screening is cheap on purpose: it ruled out 835 candidates for Hand & Stone before any deep research was spent.
Can we change what matters?
Yes. The score is plain code with the rubric written out. Change a weight, rerun the score, and every record reranks in seconds without touching the research.
What can it not find?
What search cannot reach. A state list published as a PDF behind a form will not be fully harvested in one pass. We say so in the handoff rather than pad the list.
Bring us the project that matters most.
Tell us what you’re building and what’s in the way. You’ll hear back within 24 hours.