Home Work About Blog Now Uses Contact

Alberto Rosas

Eval-first AI engineering.

> boot alberto-rosas · Mexico · since 2014

[ OK ] running

$ agent run --case analytical --eval ragas

  • router intent → analytical_query
  • retrieve query_points · prefetch ×2
  • fuse RRF · bm25 + modernbert-768d
  • rerank cross-encoder · 24 → 6
  • graph neo4j · 2-hop traversal
  • generate sql · 7 joins · 2 CTEs
  • grade faithfulness ✓ answer_relevancy ✓
  • trace langfuse
SADIE · NL2SQL accuracy 93–97%
  • router intent → analytical_query
  • retrieve query_points · prefetch ×2
  • generate sql · 3 joins
  • grade faithfulness ✗ · context recall low
  • rewrite decompose · narrow scope
  • retrieve +grounding docs
  • generate sql · rewritten
  • grade faithfulness ✓ answer_relevancy ✓
  • trace langfuse · added to eval set
CRAG · hallucination grading caught ✓
  • router intent → tool_abuse · flagged
  • guard prompt injection detected
  • mcp write op requested · denied
  • policy read-only boundary · enforced
  • respond refusal · audit note attached
  • trace langfuse · security event
MCP boundary · policy-enforced denied ✓
role
AI & Agentic Systems Engineer
exp
12+ yrs software / 5+ yrs production AI
base
Mexico · EN/ES native
Alberto Rosas
Let's talk

how i help

I help mid-market teams put AI to work: finding where it pays off, building the systems, and keeping them reliable in production.

Advisory

Point AI at the right problem

Where AI pays off in your business, what to build first, and what to leave alone.

A first engagement recovered 70+ hours a month, about $42K a year, for an operations team.

Build

Production agentic systems

Design and build the system end to end: agents, retrieval, tool-calling, and the evals that gate what ships.

Review

Second-opinion on production agents

Context engineering, evals, MCP attack surface, and agent security.

the engineer

The Engineer

Production agentic systems for customer-facing products.

12+ years in software, 5+ in production AI. Agentic workflows, tool-calling, retrieval pipelines, NL2SQL engines. Judged on eval scores and uptime.

Earlier years in cybersecurity automation (SIEM, IDS/IPS, SOAR) shape how I design for safety in everything I ship now.

Agent orchestrationTool-calling Retrieval pipelinesKnowledge graphs Vector searchEvaluation ObservabilityProduction APIs

12+

Years in software

5+

Years in production AI

500K+

Records across production systems

93–97%

NL2SQL accuracy (SADIE)

selected work

All work

Unified Compliance Framework · 2025–2026

AI engine for a compliance SaaS, over 500K+ regulated records.

Storage360 · SADIE · 2024–2025

Production NL-to-SQL platform at 93–97% accuracy.

Incfile / Bizee · 2022–2026

Multi-agent document platform at 1000+ requests a day.

experiments

Python

Planeops

Reproduce, observe, and govern a personal AI setup as three planes.

AI Tools CLI Control Plane

TypeScript · Python · JavaScript

Claude Code Remote

Steer and approve headless Claude Code sessions from your phone, over an end-to-end-encrypted tailnet.

AI Agents Claude Code CLI

Python

Soundmind

A sound-based mind programming system that combines psychoacoustic masking, brainwave entrainment, NLP-patterned affirmations, and targeted memory reactivation (TMR) into layered audio sessions designed for behavioral change and accelerated learning.

writing

Governing the pile: a control plane for a personal AI setup

Most People Use AI to Produce. I Use It to Think.

All 2 posts

newsletter

Notes from building production AI.

Evals, retrieval, agent security, and what actually holds up in production.

Occasional notes on production AI. Unsubscribe anytime.

available for

Ways to work together.

Based in Mexico, US-timezone friendly. Project-based and retainer engagements. Direct line: alberto.rsesc@protonmail.com