AI Software Engineer · MCP Server Design · LLM Agentic Systems
AI engineer and solo founder focused on replacing manual fiscal/accounting work with production-grade agent systems. Shipped a 43-tool MCP server, multi-tier Claude orchestration, specialized agents and a full fiscal compliance SaaS; all from zero in TypeScript with strict Zod schema validations. Communicates and documents in English-first environments as a default.
Experience
Founder & Lead AI Engineer
·
DISAI
Jun 2025 – Present
- 43-Tool MCP Server: Architected and deployed a MCP server (TypeScript/Node.js, Railway) covering all three primitives — Tools for CFDI 4.0 fiscal ops, Resources for live SAT catalog injection (batch-loaded per agent domain), SSE + HTTP Streamable dual transport, multi-tenant CSD management. EFOS/EDOS blacklist sync, PAC stamping. 100k+ invoices/min at <2ms latency. MCPTypeScriptRailway
- 3-Tier Agentic Orchestration: Haiku router (classifies 10 fiscal domains + query intent in ~100ms) → 10 domain DomainAgents (Sonnet, Anthropic native tool_use, real agentic loops up to 6 iterations, parallel tool calls, self-correction on failure) → Expert Registry with scoped tool allowlists (5–8/domain) + live MCP Resource injection. Zero hallucinations. Claude APITool UseMulti-Agent
- Vector DB & MCP Resource Injection: pgvector (PostgreSQL 16) for semantic SAT catalog search. MCP Resources pre-loaded into each agent's context window before the first LLM call, eliminating hallucinated catalog codes without RAG query overhead. pgvectorPostgreSQL
- Self-Hosted AI Infra: Deployed self-hosted n8n with direct MCP client integration for SAT tool execution. Replaced Zapier/Make — 80% cost reduction, 400+ integrations, full data ownership. Deployed MVPs to Railway before scaling. n8nDocker
- Streaming React Frontend: Next.js 16 + shadcn/ui chat interface with SSE event pipeline — Activity Log, Thinking Steps, Skeleton bubbles for 30–60s async LLM waits. Compliance dashboard, EFOS monitor, CSD status, driver.js onboarding, per-call savings tracking. Next.jsReact 19Tailwind
- Langfuse Observability: Instrumented the full orchestrator with Langfuse tracing — every LLM call, tool invocation, model selection, token cost, and streaming event tracked end-to-end. Supports cost attribution per session and failure triage without log digging. LangfuseTracing
- Human-in-the-Loop Reviewer Surfaces: Built compliance dashboard and EFOS monitor as Next.js reviewer interfaces with one-click approval flows, MCP-apps for interactive responses in chat and per-call savings tracking — wired directly into the agent pipeline as the human approval layer. Next.jsHITLMCP-UI
Core Contributor / Lead Developer
·
RetailDAO
Jun 2024 – Jan 2026
- Financial Analytics Terminal: Full-stack "Bloomberg-style" dashboard aggregating real-time data from Federal Reserve, on-chain APIs and crypto exchanges via WebSockets. Rate-limiting, memory caching, and graceful degradation for a self-healing low-latency system. ReactWebSocketsAPIs
- Smart Contract Audits & Documentation: Led internal audits and built a token distribution visualization platform. Mitigated investor risk and drove a 30% increase in token valuation to all-time highs.
Crypto Research Analyst
·
Coinsider
Mar 2024 – Sep 2025
- Data-Driven Intelligence: Delivered 25+ peer-reviewed analytics reports (1,200+ pages) — SWOT analysis, investor risk models, and macroeconomic forecasting for Web3 ecosystems. Translated dense technical data into content generating +70k views.