Retrieval, agents, and the scaffolding around them — embedding pipelines, tool protocols, and the telemetry that tells you when a model has started being confidently wrong.
Answers are generated by an AI model from this site's published notes. How that is handled
Retrieval that can cite itself — hybrid vector and keyword search, fused and reranked
Agent surfaces with typed tools, scoped tokens, and an audit trail behind every write
Tracing across the whole model path, so a quiet regression shows up as a number
01
Featured · Agent Observability
session-viz
Measure your agents, not your people.
Fleet-level observability for autonomous coding agents. It parses Claude Code, Codex and Cursor transcripts locally to show what agents actually delivered, surface the runs that finished having produced nothing, and decompose spend by token class — keyed on task and agent version, never on individuals.
A free, open-source Cardano-native multi-signature wallet for teams and DAOs — M-of-N approval on every transaction, collective voting on on-chain governance, DRep registration, and a documented REST API with an agent skill so a bot can co-sign under exactly the same quorum a person does.
A live visual debugger for LLM internals — it streams per-token, per-layer activation stats (attention heads, MoE routing, logit-lens flips) from a hooked model into an explorable dashboard, and runs curated small models right in the browser on WebGPU.
I care about grounding, latency, and failure modes. AI systems rarely break loudly; they return something plausible and wrong, and keep doing it until somebody checks.
Ground It
Retrieval before generation. An answer that cannot point at its source is a guess with better prose.
Constrain The Surface
Typed tools, explicit schemas, scoped credentials. A model should not be able to reach anything it was never granted.
Instrument Everything
Spans per pipeline stage, token counts, grounding checks. The failures worth fearing are the ones that never throw.
Leave A Handle
Escalation to a human, consent before an agent acts, an audit row after. Autonomy is fine; unaccountable autonomy is not.
A quick snapshot of public repository output and activity.
Public Repos
Total Stars
Featured Repos
Latest Push
03 / Work
Project Timeline
A live view of public repositories — first as a timeline of when each was built and last active, then as full detail cards.
Live · Even Realities G2
What I see
A heads-up display on a pair of smart glasses, wired to this site’s own knowledge base. The panel below is not a screenshot — it is the same 576×288 frame the glasses draw, at the five brightness steps that display actually has, filled from the same endpoint they fetch. Drive it.
TAP CHANNEL · SCROLL CARD · DOUBLE-TAP TO CLOSE
—
On the glasses these are two temple gestures: tap changes channel, scroll changes card. Same frame, same five brightness steps, same data.
04 / Contact
Let's Build
Open a secure thread or reach out directly to collaborate on AI systems and platform work.
