Building sharp digital products with engineering depth and AI precision.

Multi-tenant B2B SaaS for the Brazilian real-estate market, architected end to end (Next.js App Router, tRPC, Prisma, PostgreSQL). I designed a LangGraph + Claude agent layer spanning 11 task types for lead analysis, response generation, and campaign automation.

Causal Rationale Attribution Faithfulness Test — an open-source library that empirically checks whether an LLM agent’s stated reasoning actually drove its decision, using a randomized perturbation experiment with a placebo control. Zero runtime deps, mypy-clean, CI-ready CLI.

Adversarial stress-test harness that injects YAML-defined attacks into tool-using LLM agents and uses a Claude LLM-judge to classify compromise vs. safe refusal vs. detection. Covers 3 victim models, 5 attack categories, and 4 safety metrics, persisted via Prisma with a Next.js dashboard.

Production CLI/library that runs AI coding agents (Claude Code, Aider, GPT-4o, Gemini) inside isolated Docker containers backed by git worktrees, then fast-forward merges the results. asyncio + Pydantic v2 + Typer, 5 backends, 48 passing tests.

Multi-agent SIEM testing whether LangGraph-orchestrated agents beat a monolithic LLM on threat-detection accuracy and cost. Ingests real intrusion datasets (CIC-IDS2017, UNSW-NB15), streams events over Redis, and gates response actions behind a tRPC human-approval UI.