I architect LLM systems that cut inference costs 12x

Built a custom multi-agent runtime from zero: an LLM-driven state machine that reads intent and journey stage on every message with embedding-based intent classification, injects only that step’s prompt and MCP tools, keeps cross-session memory, and routes each step to a cheaper or smarter model — GPT-4o mini, Claude Opus, Gemini Flash Lite, Grok 4, Llama 3.1 8B, Qwen or DeepSeek, among 20+ providers. Structured outputs, guardrails, prompt caching, streaming, fallbacks and spend caps keep it at $0.02 per conversation instead of $0.25 (12x) with sub-second p95 turns, tracked on per-client cost dashboards. Retrieval is hybrid multimodal RAG over client catalogs — LLM-aware chunking, BM25 plus Vertex AI multimodal embeddings on Upstash Vector, and GCP reranking — so product search covers text and images, with GCP message queues wiring the pipelines. It handles 75% of customer conversations without a human for 15 paying B2B clients, and an eval harness took the pass rate from 43% to 87%.

• TypeScript
• Python
• LLM Routing
• RAG

Built a TypeScript library that enforces hard spending and usage caps across multiple AI providers — preventing budget overruns before they happen. Supports automatic failover to backup models when limits are hit and recycles unused capacity in real time. Drop-in installation, no servers required. Currently in production at Closer AI and open-sourced.

• TypeScript
• Node.js
• Redis
• LLM APIs

Built a zero-configuration TypeScript library that converts raw AI output into properly formatted WhatsApp messages — handling lists, product cards, links, and Spanish punctuation automatically. One function call replaces days of custom formatting work per integration.

• TypeScript
• Node.js
• NLP
• WhatsApp

Built a multi-tenant platform for building agent-powered SaaS products: a visual agent builder in Next.js and React, Node.js services on Postgres and Redis, and event-driven scheduled jobs. Every pull request runs format, lint, typecheck and 225 Jest test files in CI. Open source and separate from Closer AI.

• Next.js
• Node.js
• Postgres
• Redis

Architected the client-side IDE for a low-code platform from the ground up, including an in-browser bundler and compiler that fetched npm dependencies inside the browser and showed the real React output in the editor, kept fast with WebAssembly, caching and regression tests. The fully client-side architecture took the platform to 1M+ users while keeping infrastructure spend near zero.

• React
• TypeScript
• WebAssembly
• Compiler

Led a 10-person engineering team building MercadoLibre’s Affiliates program on Java/Spring Boot microservices through its full growth phase — from first user to 800K. The program grew to contribute 1% of MercadoLibre’s global revenue (attributed). Owned system design under real growth pressure: queues and async jobs for peak load, observability, reliability thresholds. Zero SLA breaches across 16 months of 800x growth.

• Java
• Spring Boot
• MySQL
• React

Built a Python-to-C transpiler that opened POS terminal development from a small C-specialist team to the company’s broader developer base. Deployed across thousands of payment terminals nationwide. The transpiler removed a structural hiring and knowledge bottleneck that had been limiting the company’s hardware roadmap for years.

• Python
• C
• Embedded
• Jenkins

Identified that engineers were rebuilding identical scaffolding from scratch on every client engagement. Designed and built an automated Kubernetes-based platform that cut setup time in half and standardized delivery quality across the full engineering team.

• Kubernetes
• Docker
• Node.js
• CI/CD