in production
The Careful Machine
Evidence-heavy research AI that verifies every claim against its records before anything surfaces, and declines when support is thin.
retrievereasonverify evidenceciterefuse
Miguel Sánchez Durán— senior software engineer · AI systems · Dubai
I build AI you can actually audit.
buildingsecure ai systems, rag + evidence, ai eval harnesses, full-stack platforms, agentic workflows
AI you can put in front of real decisions: it reasons over retrieved sources, shows its working, and knows when to decline.
- 15+ years engineering
- Angular / Node / Python
- RAG + evals
- local-first AI
- Dubai
Proof
Answers when the evidence supports it. Refusals when it doesn’t.
“Who approved invoice #4417, and under what authority?”
- retrieve evidence
- requirements covered
- cite
every claim ships with its records
“Had the vendor supplied them before 2024?”
- retrieve evidence
- pre-2024 history missing
- refuse
the gap is named, nothing is guessed
Systems
Three systems, in the field.
Production systems I architected and shipped, built to hold up under real constraints.
architectural views only · sensitive implementation details intentionally omitted
internal tool
The Training Forge
A gated, human-audited loop that improves local models only when a candidate provably beats the version it replaces.
correctioncandidateimmunity gateevaluatepromoterollback
daily driver
The Build Machine
An agentic toolchain that lets one engineer change large codebases at speed, with review and verification built into the path.
planchangetestinspectmergerevise
Open source
Tools I've published.
Small standalone packages, zero dependencies, one job each, with tests and CI. The reliability and control tools came out of one body of production work; the evaluation utilities are fresh implementations of standard methods. Published together in 2026, eleven of them on npm. Each one carries a CLAIMS.md mapping every falsifiable claim on its README to the test that enforces it.
A local dashboard for Claude Code: sessions, events, cost, worktrees and reviews in one loopback web app. Zero dependencies, no build step, and a real security boundary.
npx clawdeck-panel run --checkout /path/to/your/project · every release tagged, CI attached · issues welcome
The reference implementation of the pattern the tools share: the model proposes, deterministic code certifies, and every design rule is a test you can run.
five recorded live runs, claude-sonnet-5, captured 2026-08-31 · the live bench runs the published verifier in your browser
Evidence & verification
ML evaluation & calibration
Runtime & process controls
Install
npm install @m-sanchez/<name>
Eleven of the tools are published npm packages under the @m-sanchez scope, and gpu-quiescence is on PyPI (pip install gpu-quiescence). The reference implementation and the study install from a pinned git tag.
Four of the tools in one reproducible experiment: route to specialists, freeze the bars, test the win, check the confidence. In the designed world routing clears the accuracy bar and fails calibration. Against a real model it simply lost, by 15 points, and the same instruments said so.
synthetic, seeded, every number pinned by test · real-model arm recorded 2026-09-02, 400 questions, claude-haiku-4-5 · routing lost
git clone https://github.com/m-sanchez/routing-study && cd routing-study && npm ci && npm test
These are weeks old. If one of them earns a place in your build, or fails to, an issue on the repo is the feedback I most want.
Full profile on GitHubExperience
Fifteen years. Recent work first.
Weighted to current AI and full-stack work; earlier roles in brief.
Senior Software Engineer, AI / full-stack
- Evidence-heavy AI workflows end to end: retrieval and extraction that turn multi-source data into cited, reviewable output.
- Orchestration and promotion gates that keep a person in control, with agentic tooling speeding delivery.
Senior JavaScript Engineer
- Full-stack platforms for government, healthtech, and e-commerce in regulated, sensitive-data environments.
- Early NLP, retrieval pipelines, and LLM-assisted features as the tooling matured.
Earlier engineering roles
Education
MSc, Web & Mobile Application Development · 2011–2012
BSc, Computer Software Engineering · 2007–2011
Universidad de Sevilla · Seville, Spain
Writing
Notes from building careful AI.
What 15 Years of Software Engineering Taught Me About AI Engineering
Which classical engineering disciplines survive contact with stochastic systems, and which problems are genuinely new.
Building Reliable Agentic AI Systems
The Silent Zero: Proving Absence in Very Large Data
How I work
Full-stack, end to end
Frontend, backend, data, and the glue between them.
Evidence before confidence
Retrieved facts, citations, and refusal when support is weak.
Careful by default
Local-first, controlled, and reviewable, with a person signing off.
AI-augmented delivery
Agents accelerate the work without owning the decision.
Tools
- Frontend
- AngularTypeScriptReactRxJSNgRxD3 / Leaflet
- Backend
- Node.jsPythonJavaGraphQLRESTSQL / MongoDB
- AI systems
- LLM orchestrationRAG · ReActVector searchONNX / NERMCP / agents
- Delivery
- CI/CDDockerQuality hooksSecure / air-gapped
Let's build something real.
Always happy to talk AI engineering, compare notes, or dig into a hard problem.