Lucas Rosati lucasrosati

Lucas Rosati

Engineering with LLMs, not just prompting them.

SitecomAI · AI Engineer. Multi-provider AI stack with capability-based routing, provider health monitoring, cost tracking and deterministic CI replay, plus agent runtimes and CI infrastructure for autonomous coding.
MedFoco · Founder/CTO. SaaS for Brazilian medical education: AI-generated study content for students from P1 to P12 (React, PostgreSQL, Claude API).
GNI · Global News Intelligence · CTO. Predictive intelligence platform: multi-source news monitoring, scenario forecasting, risk boards and automated intelligence reports.
Argos · Agentic-first personal ERP. Agents operate it through a remote MCP server; tasks close themselves on PR merge.
FORGE · Co-founder.

Multi-agent orchestration: subagents, git worktrees, parallel Claude Code and Codex CLI sessions
Persistent memory for coding agents (Obsidian Zettelkasten as long-term context)
Context engineering: AST-derived codebase knowledge graphs (Graphify) as an alternative to embedding-based RAG
Model routing and cost optimization across providers
Automation pipelines that turn manual workflows into systems

claude-code-memory-setup · 1k stars · Persistent memory and codebase knowledge graphs for Claude Code (Obsidian + Graphify). Cuts tokens per session by up to 71.5x. EN and PT-BR docs.
agent-throttle · Run several coding agents on one machine without taking it down: worker ceilings enforced by hooks, memory checks and tuning metrics.
greenlight · Serial task queue for headless Claude Code. Nothing merges without a human.
claude-security-agents · Red Team and Blue Team agents for Claude Code that automate security audits.
claude-maestro · Parallel Claude Code agents via tmux, single script, no frameworks.

AI & LLM Systems

Claude API (Batch, Structured Outputs), OpenAI, Gemini, Deepgram, OpenRouter
Model gateway: capability-based routing, provider fallback, health checks, per-call cost ledger
Model tiering by task: Haiku triage, Sonnet editorial, Opus orchestration
RAG with hybrid retrieval and grounded citations (pgvector, embeddings routed through the gateway)
Schema-gated LLM output: Pydantic, JSON Schema, defensive JSON extraction that fails loud
Voice AI: real-time STT → LLM → TTS over WebSockets
Deterministic replay of LLM calls in CI

Agent Platforms

Isolated Claude Code runners with egress boundaries, lifecycle gates and session memory
Autonomous queues for headless Claude Code: SHA-bound CI gates, review babysitting, nothing merges without a human
Remote MCP servers with bearer auth, built so agents operate the system as first-class users
Multi-agent orchestration: git worktrees, Orca, Claude Code + Codex CLI in parallel, per-machine resource throttling
Persistent agent memory: Obsidian Zettelkasten, session hooks, cross-agent handoffs
Context engineering: AST-derived codebase knowledge graphs (Graphify), up to 71.5x fewer tokens per session

Backend & Distributed Systems

Python: FastAPI, SQLAlchemy 2, Alembic, Pydantic v2, asyncio, Dramatiq, APScheduler
TypeScript: NestJS, Prisma, Socket.IO, Node.js workers
Event-driven integration hubs: canonical models, anti-corruption layers, outbox pattern, idempotent consumers
Multi-tenant architecture: Postgres RLS, RBAC, immutable audit trail with hash chain
Execution state machines and cost ledgers with exact decimal money
Integrations: JIRA, Stripe, Telegram, WhatsApp (Evolution API), Google Ads/Analytics APIs, GitHub webhooks (HMAC), OFX bank statements, S3

Data

PostgreSQL 16/17 (partitioning, RLS, advisory locks, triggers), pgvector, Redis, SQLite
Ingestion pipelines: RSS and Telegram feeds, dedup and fuzzy clustering, legacy content migration
Headless CMS on Postgres (Payload + Next.js)

Infra & DevOps

Docker, Docker Compose, GHCR images, Caddy, Cloudflare
GCP Cloud Run, DigitalOcean, Linux, systemd
GitHub Actions: required checks, ruleset drift detection, agent-driven fix runners
Observability: OpenTelemetry, Prometheus, Sentry
Release safety: snapshot before deploy, rollback runbooks, backups off the host

Testing & Quality

pytest, Hypothesis (property-based), pytest-randomly, Vitest, Maestro (mobile E2E)
ruff, mypy, Biome, pip-audit
E2E tests against real databases

Frontend & Mobile