Problem

Every AI agent platform is either a toy chatbot, a developer library with no governance, or a cloud service that businesses can't trust with their data. Nobody is building the control plane that real companies need.

Build

A monorepo platform with workspace isolation, agent versioning with draft/publish workflow, policy-based tool access, local directory connectors with retrieval-augmented generation, streaming execution with full audit trails, and an OpenClaw runtime adapter for model dispatch.

Result

Six implementation sprints complete. Auth, agent CRUD, conversations, streaming, connectors, and retrieval all working. Full documentation suite with deployment and security guides.

What changed

Validated that the gap between 'agent runtime' and 'business-ready agent platform' is enormous. Governance, versioning, and connector management are the actual product — the LLM is a commodity.

Problem

Creating educational content is painfully slow. A single quality course takes weeks of writing, editing, and structuring. People want to learn practical skills in 15-minute chunks, not hour-long lectures.

Build

An end-to-end pipeline: topic → outline → lessons → slides → audio narration → published course. Custom TypeScript CLI (Skillbite Factory) handles generation. The platform has auth, payments, progress tracking, audio playback with synced slides, and an RSS feed.

Result

127 courses published covering everything from negotiation to distributed systems. ~18 hours of content. Revenue-ready with Stripe integration. Ideas backlog of 440 topics (148 built, 276 brainstormed).

What changed

Proved that AI content pipelines can produce at scale when you build quality control into the process, not after it. The bottleneck moved from creation to distribution.

Problem

I had 3+ years and 5,191 conversations with ChatGPT but no way to search them, find patterns, or understand how I actually use AI. The export is a giant JSON blob.

Build

A full analysis platform: SQLite with FTS5 for full-text search, Python Flask API with sentence embeddings for semantic search, Express frontend, topic clustering, behavioral scoring, and a psychographic portrait generator that categorizes every conversation by interaction mode.

Result

Discovered I'm a learner first (909 convos), thinker second (610), delegator third (446). Top use: message drafting (595 convos) — I use AI as a social co-processor. The karaoke rabbit hole alone spans 228 conversations. Generated 10 blog post outlines and identified the top recurring themes.

What changed

Reframed my own relationship with AI tools. Also generated the foundation for 'The Singularity Was Subtle' blog series — real data about how humans actually use AI, not speculation.

Problem

First-time founders waste months finding, vetting, and coordinating the professionals they need (designers, developers, marketers). There's no single platform that matches founders with execution capacity.

Build

A full-scale platform: ~62,000 lines of route code, ~25 database tables, 80+ pages. Desk workspace with settings panels, session simulator, reasoning panel with per-message AI insights. Dual backend — Express for the web layer, Python FastAPI + CrewAI for AI orchestration.

Result

Deployed and functional on Replit. The sheer scale of the build proved I can ship complex, multi-service platforms solo. The reasoning panel — showing AI decision-making per message — became a pattern I reuse.

What changed

Learned that marketplace dynamics are brutal for a solo founder. Two-sided supply/demand is a distribution problem, not a technical one. The technology works — the business model needs a wedge.

Problem

When I started building AI agents, the existing frameworks (LangChain, CrewAI) were either too opinionated, too leaky, or too coupled to specific models. I needed a runtime I could trust and understand completely.

Build

A ground-up TypeScript agent runtime: tool registration and dispatch, conversation memory, streaming response handling, multi-step execution with checkpoints, and model-agnostic adapter layers. 308 commits over months of iteration.

Result

A production-quality runtime that powered multiple downstream projects (Hiro, Clawback experiments). Deep understanding of agent execution patterns — tool ordering, context management, failure recovery — that no tutorial teaches.

What changed

Building a runtime from scratch is the single best way to understand what agents actually are vs. what the marketing says they are. Most 'agent frameworks' are prompt wrappers. The real work is execution reliability.