BuildLog
LiveTurns a team's commits, PRs, and releases into draft posts for X, LinkedIn, and Bluesky.
1,560 commits · last 12 mo
290 PRs · last 12 mo
40 public repos
2020 on GitHub since
What I do
01
CoreSpeed
I co-founded CoreSpeed. The current product gives cloud agents managed access to apps, built-in tools, and durable memory.
02
Systems around the model
I build OAuth and connector paths, model gateways, usage metering, Stripe billing, and the evidence packet sent to the model.
03
Memory research & open source
Lore measures the gap between retrieving the right record and sending the right evidence to the model. I also maintain Zypher Agent and Claude Max Gateway.
Stack
what I build with Languages
- TypeScript
- Python
- Rust
- Kotlin
- C / Verilog
Frontend
- React
- Next.js
- Astro
- Tailwind
Backend & runtime
- Cloudflare Workers
- Hono
- Bun
- Node.js
- Deno
AI & agents
- MCP
- Vercel AI SDK
- AG-UI
- Claude
- Codex
- Gemini
Data & payments
- Postgres
- pgvector
- Drizzle
- Supabase
- Stripe
- Railway
Selected
work
All projects → Current
CoreSpeed
Managed app access, built-in tools, and durable memory for cloud agents. I work on its gateway, billing, and public launch.
Fixo
A mobile-mechanic platform where an AI agent diagnoses faults, writes estimates, then handles scheduling and Stripe payments.
Amazon Order Wizard
An offline-first browser extension for tracking Amazon orders, with optional sync through a Rust and MongoDB API.
Claude Max Gateway
OSSA dual-format API gateway that exposes OpenAI and Anthropic Messages APIs through the Claude Code CLI.
Zypher Agent
OSSA minimal, open-source framework for building AI agents with full control over tools, providers, and execution flow.
FPGA Clock
A multi-function digital clock built in Verilog, running on an FPGA.
Approach
A model call rarely decides whether an agent works. The failure usually sits one boundary earlier: identity, tool access, state, billing, or the evidence packet.
Writing
All posts →How we took LongMemEval from 80.0 to 94.2 without touching retrieval
LongMemEval-S climbed from 80.0 to 94.2 while retrieval recall@10 stayed at 0.99. The gains came from evidence budget, reader generation, and instructions.
Read →We retrieved the memory, then dropped the answer
A 200-question conflict run scored 0.800 at the parent-Memory boundary and 0.635 at the evidence boundary. The gap was in what reached the reader.
Read →The setup tax: why connecting AI agents to apps is hard
Connecting AI agents to apps means OAuth, token refresh, scopes, rate limits, pagination, and webhooks per service. Here's why that work stalls agent projects.
Read →Get in touch