Capabilities

Two things, and they are related

The first is finding data that is public but nobody has bothered to assemble. The second is a way of working that makes assembling it affordable. Neither is impressive on its own.


Part one

Finding data with agents

Two of the projects on this site are databases built from records that were always public. The work was not getting hold of them. The work was collecting them without quietly poisoning the result.

Four rules do most of that, and every one of them costs something in the short term.

Work out where the answer lives before searching for it

Which free sources can actually answer the question, and which are paywalled dead ends dressed up as leads. Doing this first routinely halves the time everything else takes.

Permission is a gate, not an afterthought

Collection is gated on robots.txt. A source that disallows it is never touched — not throttled, not worked around. Of the first 91 chamber directories reviewed for Boomer, 43 publish a blanket disallow and have never been scraped.

That is roughly half the obvious sources given up on purpose. It is the right trade and it is worth knowing that is how we work before you hire us.

Never guess a value

If a field is not found, it stays empty. A wrong phone number is worse than a missing one — the missing one costs you a minute, the wrong one costs you a call, a bad impression, and the belief that the rest of the file is fine.

Provenance is stored alongside the data, not remembered. In AdvisorSearch, data_lineage and processing_log mean any single value can be traced back to the filing it came from, months later.

A human keeps the irreversible decisions

When two records look like the same business, the system proposes the merge and stops. It never merges on its own.

This one is deliberate rather than unfinished. A wrong merge fails silently — a business vanishes from the pipeline and nobody ever notices it is gone, because there is no gap to see. Automating that step would save minutes and cost trust.

What we can source

These are sources already used in live work, not a wish list.

Federal filings

SEC Investment Adviser records — the full firm feed, the monthly ADV filing data, the individual adviser feed, and ADV Part 2 brochures parsed for succession language.

State records

Business registration records and professional licence files — primary sources, which is what makes a list verified rather than merely long.

Directories

Chamber of commerce and trade directories, limited to those whose terms permit it.

Whatever your market publishes

Most regulated industries file something publicly. If yours does, it can usually be assembled.

What we won't do

  • Scrape a source that says not to. Not slower, not through a proxy, not at all.
  • Fill a gap with a plausible guess. Empty fields ship as empty fields.
  • Merge records automatically. The one manual click is the point.
  • Resell data whose licence has not been checked. Collecting something and having the right to sell it on are two different questions, and they get answered separately.

If a project needs any of the above to work, it is not a project we can take.


Part two

Agent workflows

Software like this used to need a team. It now gets built by two crews of AI agents, each with a defined job, run against one written plan.

Nineteen roles across the two crews: two leads and seventeen specialists. Research establishes what is true and worth solving, then hands off to build. You never talk to any of them — that is the point.

The two agent crews and how work moves between them A research crew of nine specialists reporting to a research lead on the left. A labelled handoff arrow in the middle. A build crew of eight specialists reporting to a delivery lead on the right, with the builders drawn as a parallel lane and a QA loop that returns to the builders until the work meets a written definition of done. The same names are listed as text below this diagram. Research crew research-lead source-librarian research-analyst primary-researcher quant-analyst data-curator fact-checker methodology-reviewer opportunity-analyst research-editor verified findings hand off Build crew delivery-lead business-analyst solution-architect ui-ux-designer Built in parallel frontend-developer backend-developer fullstack-developer devops-engineer Screens and logic at the same time, against one plan. qa-automation bugs go back Definition of done Written before anyone builds. QA loops until it is met. Verification is structurally separate. The fact-checker re-verifies from scratch, never sees the original researcher's sources, and is never the agent who produced the work.

The same nineteen roles, as text:

The parts that actually matter

Verification is structurally separate

The fact-checker re-verifies findings from scratch, without the original researcher's sources or notes, and is never the agent that produced the work. Self-auditing cannot give you that. It is a structural check, not a diligent one.

Two different questions get asked

The fact-checker asks whether a fact is true. The methodology reviewer asks whether the evidence justifies the claim being made from it. The second failure is more common and more damaging, and almost nobody checks for it.

QA loops against something written down

Not until it looks finished — until it meets a definition of done that existed before the build started. Every requested feature works, or the exception is stated in plain English.

Cost is a design constraint

Cheaper models do the grunt work, the expensive one is saved for planning and hard problems, and everyone works from a shared plan instead of re-reading the whole project. That is why a build like this is affordable at all.

How this site was built

By that crew, in about a day.

A written spec and a content file listing every fact that was allowed on the page came first, with every number queried live from the real databases so nothing could drift. A designer role built the design system. Four builders wrote the eight interactive demos in parallel, two more wrote the case studies, and everything was checked against a blacklist of claims — no invented metrics, no fabricated testimonials, no client data — before it went anywhere.

Two real bugs were caught in review and sent back to the agents that wrote them: a theme control whose label was inverted against the system setting, and a set of tab buttons that were submitting a form instead of switching views. Both were fixed by their authors, which is how the loop is supposed to work.

Everything you are reading is the output of the process it describes. That is the cheapest proof available, and it is the honest one.

The same crew is available for your build.

See what that could look like