Open to Staff · Principal · Architect roles · Madrid · Remote EU
I'm Petru Arakiss, an AI systems engineer in Madrid. Twenty years in software, in ML since 2015, and full-time on production AI these last few years. I design and ship retrieval, agent runtimes, guardrails, and evals: the parts that decide whether a system holds up after launch, not just in a demo.
01 / background
I've built software for twenty years and worked in machine learning since 2015, well before the current wave. Before AI, I shipped for banks and retailers, where a bug cost money and a missed deadline stayed missed. That work taught me the layer around the model is what matters: retrieval quality, runtimes that behave the same way twice, guardrails, and interfaces an operator can still trust when a prediction comes out wrong.
Full-stack architecture and platform work behind financial and B2B products, before moving into production AI.
At Atlax360 I lead AI engineering across three systems: BIFROST (document intelligence), ORVIAN (a multi-tenant workflow runtime), and Polaris (an internal knowledge assistant). They run inside live financial operations; I set the retrieval, evaluation, and guardrail standards the three share, and keeping them reliable under load is the job.
I use coding agents and language models heavily in my own work, and I know where they break. The bar stays the same: safe defaults, traces you can inspect, and software that keeps running when the input is malformed, the permissions are tight, or the API is slow.
Led engineering teams and full rewrites inside regulated financial environments and high-traffic retail.
Ran cross-functional platform teams through complex shipping schedules and legacy transitions.
IBM Machine Learning Professional and Linux Foundation Node.js developer certifications.
02 / current work
Most of what I build sits behind enterprise firewalls. These three show the kind of architecture I work on: ingestion pipelines, a multi-tenant runtime, and a guardrailed assistant.
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
The patterns behind my production work, extracted into Rust and TypeScript libraries: permission boundaries, trace evidence, governed memory, and structured observability.
Deterministic policy engine for AI coding agents: maps tool calls to capabilities, evaluates YAML rules, and signs every decision in a verifiable audit log, with hard-stops that policy can't bypass.
Self-inspecting, auditable memory for AI agents: surfaces the evidence, provenance, and health behind each recall so callers can see which memory to trust, with an optional Ed25519-signed tamper-evident ledger. Local-first, Rust.
Local-first trace recorder for AI agent runs: append-only, verifiable evidence of 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 measurably improve itself through typed, layered modifications validated against operator-defined evals within strict budgets.
04 / what i work on
I design and ship the whole path for AI that has to work past the demo: retrieval index, runtime, evals, guardrails, and the operator screen. The glue between an API and a UI is the small part.
Chunking, metadata, permissions, source quality, citations, and caching, plus the unglamorous work of returning “I don't know” when the index is wrong or the document is ambiguous.
Tool boundaries, stopping conditions, traces, handoffs, evaluators, queues, and cost control. The brakes matter more than the autonomy.
Eval sets, abstain logic, and regression traces. The discipline of proving a change helped instead of assuming it did.
Permissions, audit trails, PII handling, and human-in-the-loop review for AI running inside regulated finance, where “mostly right” isn't good enough.
Repository context, executable plans, browser checks, CI guardrails, and review loops around Codex and Claude, so their output stays legible and verifiable inside a real codebase.
Next.js, TypeScript, streaming UX, and the operator screen someone uses when the model misfires. In my systems, that screen is where adoption is won or lost, long before the benchmark.
05 / how i work
High autonomy, async by default, and I own the production outcome. I'm in Madrid and work remote-first across European time zones. I'm looking for Staff, Principal, Architect, and Forward Deployed roles where production AI is core to the product, with the reliability work that implies.
06 / questions
Quick facts on stack, roles, and fit. The CV has the full timeline.
Download CV07
I'm looking for teams that treat language models as product infrastructure, where retrieval quality, cost, latency, and safety are engineering problems with owners.