Work
I prefer a selective archive: fewer entries, better judgment, stronger systems.
Fullstack AI SaaS · CTO
2025 — Present
Multi-tenant B2B SaaS for Brazilian real-estate teams. End-to-end architecture (Next.js, tRPC, Prisma, PostgreSQL), a LangGraph + Claude agent layer across 11 task types, dual payment gateway (Stripe + InfinitePay), and full i18n + LGPD compliance.
LLM research library · Python
2026
Causal Rationale Attribution Faithfulness Test — empirically checks whether an LLM agent’s stated reasoning causally drove its decision via a randomized perturbation experiment with a placebo control. Zero runtime deps, mypy-clean, CI-ready CLI with a --fail-under gate.
Adversarial LLM harness · TypeScript
Stress-testing framework that injects YAML-defined attacks into tool-using agents and uses a Claude LLM-judge to classify compromise vs. safe refusal vs. detection. 3 victim models, 5 attack categories, 4 safety metrics, persisted via Prisma with a Next.js dashboard.
AI coding-agent orchestrator · Python
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 · Python
Tests whether LangGraph-orchestrated agents outperform a monolithic LLM on threat-detection accuracy and cost. Ingests CIC-IDS2017 and UNSW-NB15, streams events over Redis, and gates response actions behind a sandboxed tRPC human-approval UI.
ML pipeline · Python
Reproducible heart-disease classification benchmarking 5 classifiers with Optuna tuning, MLflow tracking, and SHAP explainability. Reaches 0.93 AUC-ROC with SMOTE applied correctly inside the training fold.
Backend system design · FastAPI
2025
Enterprise inventory API with ACID-compliant stock transfers: deterministic row-locking, atomic decrements, and DB-trigger-enforced immutable audit logs. Clean Controller–Service–Repository architecture; full stack starts with one docker compose up.
Consumer product · TypeScript
Instant iPhone resale valuation with a personal device vault. Next.js (App Router), Clerk auth, tRPC, Prisma, and Tailwind, with a real-time price formula gated behind authentication.
Scientific computing · C++
High-performance N-body gravitational simulator for the solar system, built in C++20 with symplectic integrators and real NASA ephemeris data.
AI search infrastructure
2024 — Present
Search and discovery layer for sourcing high-intent opportunities from noisy, unstructured real-estate channels.
Personal library SaaS
A tiered-subscription reading platform with EPUB and PDF rendering, annotation layers (quotes, bookmarks, notes), and per-user storage accounting. Built on Next.js 15, Drizzle ORM, Neon Postgres, Cloudflare R2, Clerk, and Stripe.