AI Features
How the AI behind this site actually works — not a demo, but production retrieval, voice, and agent-accessible tooling.
Retrieval-augmented chat
The assistant answers from a curated knowledge base, not from the model's memory. Content is embedded and stored in a pgvector index; each question runs a semantic search, and the retrieved passages are fed to the model as grounding. Answers come back with citations to the source so claims can be checked rather than taken on faith.
Content is chunked, embedded, and stored as vectors in Postgres.
Each query retrieves the most relevant passages by cosine similarity.
Responses cite the retrieved sources instead of improvising.
Voice mode
Feature-flaggedA real-time voice interface built on WebRTC. The browser negotiates the audio session directly with the realtime model using a short-lived ephemeral token, so the long-lived API key never reaches the client. Transcripts stream back over a separate WebSocket channel — distinct from the audio path — for display and grounding.
Voice is gated behind a feature flag and enabled per environment; it is not on by default everywhere.
Connect your own agent (MCP)
A Model Context Protocol server exposes the same retrieval as read-only tools any MCP-capable agent can call directly. No tool invokes a chat model or mutates anything — they are pure reads over the public corpus.
searchSemantic retrieval across the CV and this site's knowledge base, with citations back to the source.
get_timelinePublic experience or project history in chronological order, most recent first.
get_cvBulk export of the public CV corpus for an agent to read in one call.
Access is scoped: the voice path may call search and get_timeline; external clients additionally get get_cv. Every call is audited and rate limited per bearer token.
Loading…
How these are kept safe — prompt-injection defenses, fail-open design, and rate limiting — is covered on the Security & Quality page.