The work your team still does by hand, turned into software.

Real problems, working systems.

Each project below shows the business problem in plain language first, then how it was solved and the engineering decisions behind it. Client work under NDA uses recreated visuals with fictional data. The decisions and outcomes are real.

Production · brand anonymized

How the system works, step by step

Permit PDFinput→Read & interpretAI→Exact routeour code→Check per statetests→Navigationdriver-ready

■ coral = deterministic code we wrote  ·  ■ green = AI, kept only where it helps

DOT Permit Routing

Trucks hauling oversize loads must follow the exact roads listed in a state-issued permit, a dense PDF no driver can navigate by. We built the system that reads those permits and turns them into turn-by-turn navigation routes drivers can trust.

The problemEvery state formats permits differently; getting the route wrong can mean a bridge strike or a fine.

Our roleEnd-to-end: parsing, routing logic, AI integration, reliability.

DecisionThe AI kept inventing wrong turns, and no prompt fixed it. Exact geometry can't come from a probabilistic model, so we wrote code that pulls the real route from the state's own QR data.

OutcomeExact routes drivers can trust, from permits no two states format alike.

TypeScript · PythonAI + deterministic codePDF parsingMapping APIs

client withheld · visuals recreated

Test environment · fictional data

How a user moves through it

Guided questionsstep by step→Rules check the caseour code→Simple estatewill created online

↳Complex estatereferred to an attorney

■ the safety logic is rules we wrote, not AI. AI only answers users' questions about the process

Estate Map

Most Texans never write a will. Lawyers are expensive and the process is intimidating. We built a guided application that walks residents through creating one online, and knows when a case is too complex for self-service and needs a real attorney.

The problemEstate planning is high-stakes: a wrong answer can't be "probably right". The product must know its own limits.

Our roleProduct logic, guided onboarding, the decision rules, attorney escalation.

OutcomeLegal complexity turned into a flow anyone can complete, with a safety net built in.

Guided workflowDecision rulesAI-assisted support

Production screenshot

AskUpstream

A premium agriculture newsletter had years of valuable analysis locked in its archive, hundreds of issues nobody could search. We built the assistant that lets subscribers ask questions in plain English and get answers grounded in that archive, with citations.

The problemSubscribers paid for expertise they couldn't find again. The archive was valuable but unusable.

Our roleFull build: content ingestion pipeline, retrieval system, subscriber auth, admin analytics.

DecisionAnswers are grounded in the archive with citations. Retrieval first, generation second, so the AI can't invent expertise.

171active users

3,691AI messages

100%archive searchable

OpenAI vector storesBeehiiv APIRAG

Claude writes a lot of our code. We decide what ships.

"AI-assisted" doesn't mean typing prompts and hoping. We build with Claude Code. Every project runs through the same eight steps. Claude handles the heavy lifting in some, and the steps that decide quality stay with us: understanding the problem, reviewing the code, and taking responsibility for the result.

Most of our projects start with an expensive manual process.

Someone is tracking inventory in spreadsheets, looking up prices by hand, or planning a school year across disconnected documents. We turn that into software that runs the operation on its own.

GAP Smart Inventory

Inventory, invoicing and tax operations automated end-to-end for a food-retail operation.

60% less stock loss · 80% less admin time

Academic Program Operations

School planning data turned into operational progress visibility for a multi-role team: what was delivered, by whom, and how it was rated.

private system · brand and names anonymized

Translated from the original where needed.