Madrid · AI systems · public notes
I build AI systems that need evidence, permissions, traces, and boring operational checks. I write about the parts that fail after the demo, and about the cost of putting work in public.
01 / start
The useful entry points are the writing and the public repos. The CV exists for formal processes, after the work has made the conversation worth having.
2006
Professional software work across product, backend, platform, and AI systems.
2015
Retrieval, agent runtimes, guardrails, evals, traces, and operator screens.
Private production work gives me constraints: regulated finance, document workflows, retrieval quality, permissions, cost, latency, and review paths. Those constraints shape what I write and what I build in public.
Public work is where the claims can be checked: repos for agent policy, memory, traces, logging, local tooling, and the surrounding harnesses that make AI output inspectable.
Notes on production AI, agent harnesses, attention, indie hacking, and the cost of publishing work that can be judged.
This year's quiet building-in-public attempt: a memory and evidence system for long-running AI work.
Private work supplies the boring parts: permissions, latency, cost, retrieval quality, review paths, and operator screens.
02 / constraints
The implementation details are private. The shape is public enough to explain the problems: document ingestion and retrieval, workflow runtime behavior, and an internal assistant with guardrails and citations.
document intelligence
Document intelligence for the documents finance actually runs on: ingestion quality gates, semantic and visual chunking, pgvector/HNSW search, caching, source-quality scoring, analytics, and explicit no-answer paths when evidence is weak.
workflow runtime
Multi-tenant AI workflow runtime: context assembly, durable memory, deterministic and cached execution tiers, queue processing, idempotency, run events, and human-review metadata when automation should stop.
internal assistant
Internal assistant that combines BIFROST retrieval with guardrails, citations, streaming UX, suggestion revalidation, and operator analytics. One surface for support, sales, and product teams.
03 / open source
Public work I can point to directly: agent policy, traces, governed memory, structured observability, local systems tooling, and developer environments.
Deterministic policy engine for AI coding agents: maps tool calls to capabilities, evaluates YAML rules, and signs allow, deny, and ask decisions in an audit log.
Self-inspecting, auditable memory for AI agents: surfaces evidence, provenance, and health behind each recall, with an optional Ed25519-signed tamper-evident ledger. Local-first, Rust.
Local-first trace recorder for AI agent runs: append-only run evidence for what the agent called, what it was allowed, and what failed, with hook ingestion for Codex/OMX harnesses.
Runtime-agnostic structured logging with automatic PII sanitization (GDPR/HIPAA/PCI-DSS) and native W3C tracing. Zero dependencies; runs on Node, Bun, Deno, Edge, and the browser.
Research harness exploring whether a coding-agent harness can improve itself through typed, layered modifications validated against operator-defined evals within strict budgets.
Zero-dependency Apple Silicon hardware monitor and thermal watchdog in Rust: CPU/GPU/ANE power, per-core load, temperatures, fans, and battery without sudo.
Open terminal environment for Ghostty on Linux: themes, presets, fonts, tmux integration, and a config switcher for a Warp-like workflow.
04 / what i work on
Most of the value sits between a source document and the person using the result. That path includes indexing, permissions, runtime state, evals, guardrails, latency, and the operator workflow for failure handling.
Chunking, metadata, permissions, source quality, citations, and caching, with explicit refusal when retrieved sources are weak, contradictory, missing, or out of scope.
Tool boundaries, stopping conditions, traces, handoffs, evaluators, queues, and cost control, defined before an agent reaches production.
Eval sets, abstain logic, and regression traces. The discipline of checking whether a change helped before treating it as an improvement.
Permissions, audit trails, PII handling, and human-in-the-loop review for AI running inside regulated finance, where a fluent wrong answer is still a defect.
Repository context, executable plans, browser checks, CI gates, and review logs around coding agents, so their output stays verifiable inside a real codebase.
Next.js, TypeScript, streaming UX, and operator screens for model failures, showing retrieved evidence, failure state, audit trail, escalation path, and the next valid action.
05 / how i work
I work async and stay accountable for retrieval quality, runtime behavior, latency, cost, permissions, and operator workflows. If there is a serious production AI problem behind the conversation, email is the shortest path.
06 / questions
Quick facts on writing, public work, current constraints, and when a direct email makes sense.
Read writing07
Best conversations start with a broken retrieval path, an agent that does too much, an eval that catches too little, or a product surface where users cannot tell why the model answered.