Work Selected work

Each build is described by the problem it had to solve, not by the tools it used.

Every measurement on these pages carries the date it was taken. None of them is a claim about what runs today. Where a build has a cost or a limit, that cost sits next to the result.

Client names: I name no client, here or anywhere. This is my own policy. It will protect your work the same way it protects theirs.
Every measurement: Carries the date I took it. I do not use the present tense about a system I cannot look at today.
No screenshots: Work boards hold real names and real addresses. A drawing shows how a system works, and a drawing cannot leak anything.
Check it yourself: The fourth build has no client in it. You can judge its design, its list of commands and its test count on their own.

01 The builds

Each build puts its cost next to its result.

Four of them are layers of one system for a field-services company. The fifth is the engagement that shipped with this site. Open any of them for how it works, what it measured, and its honest limitation.

01 Lead bridge · Records and tracking

Two CRMs that could not say when anything changed

Six lead sources. Two CRMs that could not say when anything changed. Every side effect had to happen once.

262 of 262 runs succeeded in the seven days of history the platform keepsweek ending 2026-09-11

02 Permit signal · Signal

A weekly pipeline over public data that moves every time it is published

Public data behind links that stop working on every release. Ranked into a route the field team can drive.

3 + 1 tests written before the pipeline was switched on, plus one practice run against a hand-checked week2026-09-11

03 Document search · Knowledge

Seven sets of help pages, searchable inside the assistant they already used

Seven vendors' help pages, built to one shape. Answers name their source, inside the assistant already open.

15,788 passages across the seven document sets, from 5,899 help pages2026-09-19

04 Warehouse assistant · Assistant interface

An assistant that reaches a warehouse without being handed the keys to it

An assistant connected to a named warehouse system: sixteen commands, one that writes, six checks run again at the moment of the change. No client in it.

324 tests passing against the stand-in warehouse, and the code builds clean2026-09-18

05 Design-build contractor · Engagement

A San Diego design-build contractor

Lead intake, a job pipeline, a knowledge base, and the playbook to match, set up remotely on the client's own accounts.

02 One system

Four layers of one system, not four unrelated jobs.

The four builds sit at different heights in one stack. Each one rests on the layer below it. The permit pipeline writes its results through the lead bridge. The document search sets travel over the same protocol as the warehouse assistant. You can buy any single layer. No layer needs the others to be useful.

Four layers stacked. At the base, a records and tracking layer holding one record per lead and the tracking tables. Above it a signal layer that reads public permit data and sends qualified addresses down into the records layer. Above that a knowledge layer of seven document sets. At the top an assistant layer that reads the records and searches the document sets over one protocol.

Each layer depends on the one under it. The assistant has nothing to offer without records or documents beneath it. The documents have nothing to index without written procedures or vendor help pages. The highlighted arrow marks the one place where two of these four builds touch in daily use. The permit pipeline hands its records to the bridge instead of writing to a CRM itself.

Six words used on these pages

CRM: The software a sales team keeps its customer records in.
Webhook: A message a system sends the moment something changes. Nothing has to keep asking.
Poll: Asking a system on a schedule whether anything changed. This is what you do when there is no webhook.
Duplicate guard: A check that makes a repeated message do nothing the second time. It is what stops one lead becoming two records.
Document set: One vendor's help pages, turned into plain files that a computer can search.
Protocol: An agreed way for an AI assistant to use an outside tool. The set of rules both sides follow.

03 How to read this

How to read these pages, and the question they raise.

Three of these four builds were done for a client I do not name. Nobody made me agree to that. It is my own policy, and it will protect your work the same way it protects theirs. It does cost these pages the usual things a portfolio has: a named case study, a reference call, a logo. Here is what I offer instead.

Every measurement is one I took on a stated day and will stand behind. None of them is a claim that somebody else has checked my work. Where a number covers only seven days, the page says seven days.
The public half of each problem is something you can check without me. Can a connection announce a change? Does a city office give its documents a new address every time? What does a published test actually report? Those are the hard parts of three of these builds, and anyone can check all three in about two minutes.
Build 04 has no client in it. You can judge its design, its commands, its refusal messages and its tests as engineering. That is why it sits here next to work I cannot show you.
Nothing here is a screenshot, and nothing here is a price. Work boards hold real names and addresses, so I draw the system instead of photographing it. I price work by the result, so the money conversation belongs in a proposal.
If you want proof that I can build it, rather than proof that I did build it, the quickest route is a paid first stage on your own systems. You own whatever it produces, and you can stop there.

Builds 01 to 03 were measured on 2026-09-11 from the build records. The counts in build 03 were measured again on 2026-09-19, straight from the files.
Build 04 is as of 2026-09-18. It was tested against a stand-in warehouse, not against a live account.
The published results quoted in build 03 were read from anthropic.com/news/contextual-retrieval on 2026-09-19.
No client is named on these pages. No measurement on them was taken from a client's systems in order to write them.

Want proof on your own systems?

Paid discovery runs on accounts you own, and you keep the findings whether or not we build. Thirty minutes first, no prep, no pitch.

Book an intro call