I got tired of talking about things that worked. I wanted to make them work.
Three startups later, I'm mass-converting English fluency into system design. Turns out "how do I explain this clearly" and "how do I structure this cleanly" are the same question.
|
Multi-agent workspace shipped with PersistOS. |
Agent framework built around selection pressure. |
Compact A2A format for machine-readable coordination. |
Links
- Org: https://github.com/persist-os
- The Convergence: https://github.com/persist-os/the-convergence
- Vector Native: https://github.com/persist-os/vector-native
- neural-polygraph — SAE-based hallucination detection. Spectral + geometric probes, runners, visualizations
- experiments — ML run patterns: append-only specimens, Parquet artifacts, DuckDB catalog
- 10 Things I Wish I Knew When I Started Using AI for Coding — workflow discipline: read the code, manage context, plan first
- This AI Analyzes My Entire Life and Costs $0/Month to Run — local-first “Synthesis Pool” as a personal data system
- Why Prompt Engineering Can’t Fix Hallucinations (But Neurosurgery Can) — SAE-based signals for hallucination detection
- Stop Building Chatbots — why chatbot-first architecture breaks in production
- Beyond RAG — evolving context systems, not just retrieval
- Latency & Logic — why latency + structure matter more than raw tokens
- Cursor as Agent Civilization — building coordination layers over autocomplete
Ship things
- 6 hackathon wins (AWS, Google Cloud, Agno, Wordware, etc.)
- Each one: 24-48 hours, working code, judges who said yes
- Some became real products
Debug things
- Agent orchestration that fails silently
- Context windows that overflow
- Pipelines that work on my machine
Learn things fast
- Zero Swift → TestFlight app in 2 months
- Zero agent experience → production multi-agent system in 4 months
- Pattern: find the hard part, sit in it until it clicks
Production-minded
- Packaging, CLIs, setup verification
- Docs that assume zero context
- "Runs on a fresh machine" as a requirement
Research-capable
- Experiments that leave queryable outputs
- Metrics you can regenerate
- Minimal magic, maximal traceability
Python · FastAPI · Redis · Convex · Polars
OpenAI · Anthropic · Agno
Next.js · TypeScript · SvelteKit
GCP · AWS Bedrock · VercelSan Francisco
