I ship working AI apps by directing coding agents.
I don't hand-write code. I direct AI coding agents and verify what they actually built, so the work that ships is the work that was asked for. Everything below is live, built that way, and in use.
If you run an agency or a small team with more client builds than hands, I can take the overflow: chat assistants, lead capture and booking, small SaaS, internal tools.
Click them. They work.
Tinker: crew time clock SaaS
Multi-tenant time tracking for small crews. Employees clock in from a shared tablet or their own phone, punches get a location tag, owners get pay-period hour reports. Self-serve signup, subscription billing, per-business data walls. Its first business uses it to track real crew hours.
Lead capture + booking assistant
The chat assistant on precisionguessworktech.com. Answers only from the business's real information, says so when it doesn't know, takes leads and books appointments, and alerts the owner. Built multi-tenant so it can be cloned per client.
Trade demos: salon, restaurant, auto, home
Four ready-to-sell chat assistants for local businesses on one shared backend. Swapping the business info is the only change per client.
Document Q&A (RAG)
Ask questions about a set of documents and get answers grounded in them or an honest "it isn't in there." The standard retrieval setup clients ask for, done so it doesn't make things up.
Free hosting: the first load can take up to a minute to wake up.
Branded chat assistant
A clean, re-skinnable Claude chat app with tool use (live weather lookup as the example). The starting point for a client-branded assistant.
Native apps
Several native Android apps (a voice-controlled kitchen timer, a supplement tracker, a recipe app with photo scanning), and a Godot mobile game.
Agents that don't fumble the build
The usual problem with AI coding agents isn't bad code. It's an agent reporting work it didn't do, and you finding out weeks later. I've spent a year building a working method against exactly that, and wrote it up in public:
Directing Agents You Can't Audit: a field-tested protocol for building software with AI agents when you can't read the diff.
- Every piece of work is checked against what it was supposed to do, not against the agent's report.
- Small, testable steps with a written record of what landed and what's still owed.
- Fast: a working first version in days, not weeks.
- Plain-English updates. You always know what's done and what isn't.
Got more builds than hands?
Hourly or per project, fully remote, white-label is fine. Tell me what you need built and when.
lesfleming@precisionguessworktech.com