AI Engineer · Production LLM & Agentic Systems on Structured Financial Data

AI engineer building production agentic systems in high-stakes financial domains. In crypto and decentralized systems since 2020; since 2025 in Mexican fiscal compliance, where a wrong output is a legal filing with the tax authority. Shipped solo: a production MCP server (46 tools), a three-tier Claude orchestration layer with cost-aware model routing, and a full AI-native compliance platform (in private beta) with chat interface, STT input, CI-gated evals, end-to-end observability, multi-tenant security, human-in-the-loop approval on irreversible operations, and 8,900+ tests across two repos.

Previously led development at RetailDAO, where I ran smart-contract audits and executed the on-chain governance action that permanently capped the token supply, removing inflation and honeypot risk for holders. Python and TypeScript/Node. Async collaboration with North American teams; English-first documentation.

Experience
AI Engineer & Founder · DISAI (independent)
Jun 2025 – Present
  • Production MCP server (46 tools). Designed, built and operate a Model Context Protocol server (TypeScript/Node, Railway) covering the full CFDI 4.0 lifecycle: invoicing, payroll, carta porte, payment complements, live SAT catalog injection, EFOS blacklist sync, PAC stamping, multi-tenant credential handling. Migrated to MCP SDK v2 and the stateless 2026-07-28 protocol within a week of its release, keeping 2025-era clients working from the same endpoint. MCPTypeScriptRailway
  • Three-tier agentic orchestration on Claude. Haiku 4.5 router with a deterministic regex tie-break across 10 fiscal domains, 10 domain agents using native tool use with agentic loops and self-correction, server-synced tool allowlists, and cost-aware executioner routing (Sonnet 4.6 standard, Sonnet 5 advanced, Opus 4.8 escalation). Claude APITool UseMulti-Agent
  • Reliability under irreversible operations. Human-in-the-loop confirmation gate before any tax-authority stamping; exactly-once PAC stamping via in-process single-flight plus an atomic PostgreSQL claim; fail-open vs fail-closed policy per subsystem; DB-persisted circuit breakers on stamping (PAC) providers. HITLIdempotencyCircuit Breakers
  • Multi-tenant security. PostgreSQL row-level security, short-lived HMAC-signed assertions binding each org to the tool it calls, encrypted per-tenant signing certificates, TOTP MFA, Redis-backed rate limiting, timing-safe key checks and a structured auth audit trail. RLSHMACMFA
  • Retrieval and context injection. Domain-scoped retrieval over 95 MCP resources (SAT catalogs, templates, regulatory docs, versioned domain skills, blacklists) with live per-domain context injection. Accent-insensitive fuzzy search across 52k+ catalog entries. Designed the hybrid upgrade (trigram + embeddings, RRF) so retrieval only proposes candidates and exact validation alone commits a code, with exact kNN over ANN to keep results reproducible. MCP ResourcesRetrieval
  • Evals & observability. Routing eval harness (golden set, 4 scorers) gating CI offline on every PR and live on main; Langfuse tracing with per-session cost attribution, Sentry for errors. 8,900+ tests across two repos from unit and contract through VCR replay, integration, smoke and Playwright E2E; the MCP server holds 97% line / 90% branch coverage, enforced in CI. EvalsLangfuseVitest
  • Production ownership. 28 architecture decision records and 8 written postmortems (cross-tenant credential fallback, a bypassable rate limiter, untraced PAC routes), each tracing root cause and blast radius to a fix and a generalizable lesson.
  • Full-stack. Next.js 16 / React 19 streaming frontend, Node backend, PostgreSQL/Redis data layer, Zod 4 validation at every boundary with compiled hot-path schemas (3-7x validation throughput), GitHub Actions CI. Next.jsReact 19Zod
  • Applied mathematics in production. Directed supplier-network risk graph with BFS propagation and DFS cycle detection for tax-fraud carousel patterns; derived and published the decay bound, released as the open-source efos-risk-graph npm package (v0.3.0). The network is built from each org's own invoices and shown as an interactive graph with what-if exclusion. Graph Algorithmsnpm
  • Customer-facing delivery. Ran discovery sessions with Mexican accounting firms and translated ambiguous fiscal and regulatory requirements into working agent workflows.
  • AI-assisted engineering. Built the platform solo directing Claude Code, Gemini CLI, OpenCode and other harnesses daily, reviewing all generated code for logic, security and performance defects as the last line of defense before production.
Experience (cont.)
Core Contributor / Lead Developer · RetailDAO (Base mainnet)
Jun 2024 – Jan 2026
  • Protocol security, audit and irreversible mitigation. After external scanners flagged $RETAIL for an active mint function and opaque ownership, I led the audit that traced the root cause to the Aragon IVotes plugin retaining execute() permissions on the original mint DAO, compounded by non-standard ownership (no renounceOwnership path) and proxy contracts obscuring control. Designed and executed the mitigation: permanently revoke all execute() permissions, rendering the mint DAO immutable and capping supply at 1B tokens. Passed by on-chain token-weighted vote; external risk flags cleared post-execution. Proposal on Base mainnet · Audit and mitigation report SolidityAragonGovernance
  • Real-time analytics terminal. Built and shipped a Bloomberg-style crypto terminal aggregating 8+ sources (Federal Reserve, on-chain, exchanges, intelligence firms) over dual WebSockets, with Redis caching, rate limiting, exponential backoff with jitter, golden-dataset fallback and graceful degradation. ReactNode.jsRedisDocker
  • Protocol documentation platform. Authored and shipped the complete documentation site solo: tokenomics, audits, governance structure, treasury and budget policy, onboarding, plus the GitHub org infrastructure behind it. retaildao.github.io/Docs DocusaurusGitHub Pages
Data & Research Analyst · Coinsider
Mar 2024 – Sep 2025
  • Hired after a multi-stage technical assessment on protocol mechanics, market structure and on-chain analysis.
  • Delivered 25+ peer-reviewed analytics reports (1,200+ pages): investor-risk models, SWOT analysis, macro forecasting. Translated dense quantitative material into clear writing that reached 70k+ views.
Data Analysis Consultant · Chainlabs (short engagement)
Jun 2025
  • Delivered two wallet-analysis notebooks in Python for a European crypto-intelligence firm: OSINT techniques, cross-chain data wrangling, dataset querying, visualizations and relational graphs in Plotly and Matplotlib. PythonPlotlyMatplotlib
Foundation
2020 – 2024

Self-directed work in crypto and decentralized systems alongside a design career: protocol mechanics (PoW/PoS, L2s, zero-knowledge proofs, cross-chain), on-chain operations, multi-wallet and cold-storage management, and market structure and quantitative analysis (derivatives, leverage, risk-adjusted return metrics, systematic strategies). Externally validated by passing Coinsider's technical assessment in 2024.

Prior to that: 12+ years as a graphic designer, illustrator and tattoo artist/studio owner, before transitioning full-time into software and AI engineering.

Skills
AI / LLM
LLM application development, AI agents, multi-agent orchestration, agentic loops and self-correction, prompt engineering, system instructions, evals, hallucination control, human-in-the-loop (HITL), cost-aware model routing, retrieval and context injection over MCP resources, fuzzy/lexical and hybrid retrieval design, Claude API, tool use / function calling, MCP (tools, resources, prompts, multi round-trip confirmation, MCP Apps), CI-gated eval harnesses
Languages & Backend
Python, TypeScript, Node.js, SQL, REST APIs, object-oriented design (abstract base classes, interface-based DI, strategy pattern), multi-service architecture, streaming / server-sent events (SSE), Streamable HTTP (stateless MCP 2026-07-28)
Data & Infra
PostgreSQL, SQLite, Redis, Docker, WebSockets, rate limiting and backoff strategies, caching, circuit breakers, idempotency
Frontend
React, Next.js, Tailwind CSS, shadcn/ui
DevOps & Practices
Docker, CI/CD (GitHub Actions), Git, Railway, Vercel, automated testing (Vitest, Playwright), schema validation (Zod), observability (Langfuse, Sentry), ADRs and postmortems, technical documentation
Security
Multi-tenant isolation (PostgreSQL RLS), HMAC request signing, TOTP MFA, rate limiting, encrypted credential storage, audit logging, permissions
AI-Assisted Eng.
Claude Code, Gemini CLI, Zed, OpenCode (daily production use), AI-driven code review
Customer-Facing
Customer discovery, requirements gathering, solution delivery, UAT, stakeholder collaboration
Selected Projects

SAT-MCP Server

Production MCP server for Mexican fiscal compliance. 46 tools, 18 prompts, 95 resources, 9 in-chat MCP Apps; stateless Streamable HTTP (MCP 2026-07-28) with backward compatibility. 6,400+ tests, 97% line / 90% branch coverage.

TypeScript MCP SDK Node.js Railway PostgreSQL

DISAI-Conta

AI-native fiscal compliance platform for accounting firms, in private beta. Routing agent plus 10 domain agents, scoped allowlists, live context injection, CI-gated routing evals, supplier-network risk graph, human-in-the-loop compliance dashboard. 2,500+ tests.

Next.js 16 Claude API Tool Use Langfuse shadcn/ui

Open-Source Pattern Libraries

Reusable production patterns extracted from the platform: efos-risk-graph (graph risk propagation, BFS/DFS), agentic-tool-loop, llm-cost-router, mcp-tool-idempotency and euler-worked-out.

TypeScript npm Graph Algorithms Open Source

AI Reliability & Evaluation Lab

Reproducible benchmark for bounded tool-using agents: direct Sonnet 5 vs routed Haiku/Sonnet on fictional fiscal data. Hashed immutable datasets, versioned pricing, hard spend ceilings in CI, and reports generated only from raw run logs.

Evals Claude API TypeScript Open Source

Technical Writing

Engineering write-ups at tripl3.dev/writing: cost-aware model routing, MCP server throughput, supplier-network risk math, and field notes from building AI-native fiscal compliance solo.

Writing Architecture

n8n AI Automation Stack

Self-hosted n8n ultra-lean instance with direct MCP server/client integration and sub-agents per document type; replaced Zapier/Make for an ~80% recurring-cost reduction.

n8n MCP LLM Agents Docker
← tripl3.dev ↓ Download PDF