Around 2010, I built something called AppFrame: a framework that would take intent and produce a running application without a developer writing every screen, every validation rule, every workflow by hand. It ran in WinForms on the desktop and used .ashx HTTP handlers at the backend.
It wasn't a prototype. It took years to build, and it actually worked. You could describe what you wanted to do, and AppFrame would produce something runnable. The idea was real. The execution was real.
Then Web 2.0 arrived and rendered it almost instantly obsolete. The world moved to browsers, AJAX, and rich web UIs practically overnight. A WinForms-based app framework that generated desktop applications was suddenly building for a platform nobody wanted any more. Years of work, made irrelevant by a shift in how people expected software to feel.
So I parked it. Not abandoned — parked. The idea was right. The delivery mechanism was wrong. The tech had to catch up.
Building the missing pieces
Over the years that followed, each project I built was quietly an answer to one of the questions the original idea couldn't answer.
| Project | The question it answered |
|---|---|
| BareMetalWeb C# · .NET · metadata-driven SSR |
Can you drive a full dynamic application — forms, tables, auth, validation, workflows, storage, real-time updates — purely from a metadata description, with no framework in the way? Yes. You can get it down to a single HTTP handler and a schema definition. |
| BareMetalJsTools JS · 62 modules · zero dependencies |
Can the frontend — reactive UI, forms, charts, routing, auth, components, styles — exist without a build chain, without node_modules, without a compile step? Yes. 638 KB. Plain script tags. No committee required. |
| PicoWAL C · RP2350 · $6 microcontroller |
Can persistent storage — a real database with schema, queries, joins, aggregates, auth, and a web UI — run on $6 of hardware with no server? Yes. A database is just structured bytes on a write-ahead log. It doesn't need a committee meeting. |
| Picoweb C · TLS inline · zero proxy hops |
Can the serving layer disappear entirely — no reverse proxy, no TLS terminator, no sidecar — leaving nothing between your bytes and the wire? Yes. TLS direct to socket. Speed as a side effect of removing everything fake. |
| PlatinumForge / Forge Studio C# · .NET 10 · 8 AI agents · agentic SDLC |
Can AI actually understand intent, coach you into a proper business definition, and then build and run the application from it — the thing I was trying to do in 2010? Yes. Finally. |
What AppFrame was actually trying to do
AppFrame had the shape of the idea right. You'd describe what you were trying to do. The tool would ask you questions. You'd answer them. It would figure out the data, the screens, the rules, and the flows, and produce something runnable. You wouldn't write the app — you'd describe it, and the machine would build it. That part worked.
What it was missing — and what Web 2.0 exposed — was that the delivery target was wrong, and the intelligence layer was too thin to replace a developer's judgment. When the world moved to browsers, AppFrame had no answer. When users wanted something that felt like Gmail or Google Maps, WinForms had nothing to offer. The idea survived. The container it lived in didn't.
The other missing pieces:
- Real AI. In 2010, "AI" in this context meant hand-crafted heuristics and rule-based pattern matching. It couldn't understand vague intent. It couldn't ask the right follow-up questions. It couldn't catch missing rules or unsafe assumptions. The coaching loop didn't exist.
- The right storage primitives. Everything assumed a SQL Server or Oracle behind it. The idea needed something lighter, cheaper, and deployable anywhere. PicoWAL is the answer to that question — a decade later.
- A UI layer that didn't fight you. WinForms was the UI. Later attempts used ASP.NET WebForms and early MVC. All of them added more ceremony than they removed. BareMetalJsTools is the answer — zero-dependency, declare what you need, get on with it.
- A serving layer that stayed out of the way. The .ashx handler was the thinnest thing available then. It still ran inside IIS with all its weight. Picoweb removes the last fake steps.
- A metadata-driven app host. The prototype reimplemented the app host by hand every time. BareMetalWeb turned that into a reusable foundation: describe your schema, get your app surfaces.
Why now
Large language models changed the equation. Not because AI is magic — it isn't — but because for the first time you have something that can hold a conversation about what you actually mean, ask sensible questions, notice when a rule is missing, and produce structured output that a pipeline can act on. That's the coaching loop the 2010 prototype couldn't have.
PlatinumForge's eight AI agents — Psi, Apollo, Prometheus, Hephaestus, Themis, Hestia, Zeus, Thor — are the realisation of that loop. Not a single chatbot guessing what you want. A council that expands ideas, challenges assumptions, enforces rules, stress-tests decisions, and only lets something through when it's actually ready. Tests written before code. Constraints traceable from requirement to deployment. Apps that don't just generate once but version, evolve, and stay tied to the definition they came from.
Forge Studio is the end-user face of the same idea: describe what you want to do, get coached into a proper definition, get a running application. No developer required for the boilerplate. The developer work that remains is the work that was always worth doing.
The stack, finally assembled
| Layer | What it does |
|---|---|
| Intent + coaching | PlatinumForge / Forge Studio — AI agents turn vague requirements into a versioned, validated business definition |
| App host | BareMetalWeb — metadata-driven server-side rendering, forms, auth, storage, real-time updates, no framework |
| Frontend | BareMetalJsTools — 62 modules, reactive UI, zero dependencies, no build step |
| Storage | PicoWAL — structured bytes on a write-ahead log, deployable anywhere, no server required |
| Serving | Picoweb — TLS inline, direct to socket, zero proxy hops, bytes on wire |
Each layer removes a reason the 2010 idea couldn't ship. Together they answer the question I've been asking since WinForms: what's the minimum honest stack to go from intent to running application?
→ How Forge Studio turns a business definition into a running app
→ PlatinumForge — the agentic SDLC pipeline for developers
→ About me