hi, I'm Leonard Tan
I build low-level systems, AI infrastructure, and agents that know how to forget.
A working AI-agent company. ~11 specialist employees — sysadmin, HR, strategy, comms, three research labs — each owning a domain end-to-end. They communicate async via markdown dropped into file-based inboxes on a FUSE filesystem; no chat, no standups, no Jira. The CEO is human; the rest of the org chart isn't.
Touch ID-gated encrypted files. Stops AI editors (Cursor, Claude Code) from reading your .env and credentials — encryption happens below the application layer, so every tool just works without plugins.
Minimal agent with forgetting. Old context gets compacted — details fade but the gist remains, and the original is always one UNCOMPACT away. Content-addressed blobs, per-identity event logs, self-editable compaction prompts.
Namespaced multi-provider LLM observability proxy. One endpoint per provider (Anthropic, OpenAI, Gemini, xAI, DeepSeek, Kimi, OpenRouter). Every request persisted to SQLite with full-text search, latency, tokens, redacted secrets.
On-device speech-to-text for iOS via whisper.cpp. No network, no cloud — transcription runs entirely on the phone.