Tiago Fortunato
AI Application & Product Engineer

WhatsApp-first scheduling SaaS for Brazilian freelance professionals (therapists, personal trainers, salons), with an AI intake layer that turns natural-language WhatsApp messages into bookings. Live product at odys.com.br, currently in active validation with industry professionals before scaling marketing.

• AI intake (Groq LLM): parses Portuguese WhatsApp messages, extracts intent as Zod-validated JSON, proposes slots and books; custom prompt design, tool-calling, retry logic, and deterministic fallback built from scratch
• Multi-tenant architecture: each professional gets a public booking page at /p/[slug] with isolated data; plan-gated features (free/basic/pro/premium) with trial flow
• Stripe + PIX with idempotent webhook handling; Supabase Auth with email verification; Drizzle ORM on PostgreSQL; resolved PgBouncer transaction mode compatibility
• WhatsApp-native reminders via self-hosted Evolution API v2 on Railway, orchestrated by Supabase pg_cron with signature-verified webhooks
• Production maturity: rate limiting (Upstash Redis), Resend SMTP, Sentry, PostHog, type safety (TypeScript + Zod), CI/CD (GitHub Actions)

Production RAG chatbot where recruiters and hiring managers ask questions about my background, projects, and skills, and get answers grounded in a curated knowledge base. Migrated from LangChain to LangGraph. Self-hosted on AWS EC2.

• LangGraph StateGraph with 7 async nodes, conditional edges, and typed state: query enrichment, project routing, hybrid retrieve, context formatting, answer generation, follow-ups, plus a fallback path for empty results
• Hybrid retrieval: semantic search + BM25 fused with Reciprocal Rank Fusion (RRF) for higher precision than either method alone
• Custom section-aware chunking via font-size analysis: keeps semantically related content together for better retrieval quality
• Streaming SSE responses via astream_events with Groq (Llama 3.1), conversation history, and suggested follow-up questions
• RAGAs evaluation pipeline measuring faithfulness, answer relevance, context precision, and context recall
• FastAPI backend with ChromaDB vector store; vanilla JS frontend; containerized with Docker; self-hosted on AWS EC2 with Nginx reverse proxy and Let's Encrypt HTTPS

Production REST API for infrastructure inspections with autonomous Vision AI classification. Upload a photo of road damage and the system classifies severity, damage type, and generates an explainable rationale, all running in the background.

• Vision AI classification via Groq SDK: images are compressed client-side, sent to the API, and classified autonomously using background tasks
• Explainable AI (XAI): every classification includes a human-readable rationale field so decisions are transparent and auditable
• LangSmith integration for full LLM observability: trace every prompt, response, and latency in production
• Structured LLM output enforcement with LangChain for consistent severity and damage type responses
• JWT auth with admin role, per-user data isolation, status lifecycle, rate limiting
• Comprehensive Pytest suite covering all endpoints; CI/CD via GitHub Actions; Docker; frontend dashboard on Vercel with image upload and AI status badges

Two-stage hybrid pipeline combining YOLOv8-based object detection with a rule-based expert system to detect road surface damage and automatically prioritize maintenance actions. Trained and evaluated on the RDD2022 benchmark dataset.

• Conducted 4 controlled experiments (EXP1 to EXP4) varying model capacity, input resolution, and training length. Best config (YOLOv8s): mAP50 0.663, Precision 0.694, Recall 0.604
• Built a rule-based expert system with geometric filtering, continuous severity scoring, and quantile-based prioritization (LOW / MEDIUM / HIGH)
• Post-processing reduced noisy detections by 31.2% (8304 → 5711) while preserving structurally relevant defects
• Designed for interpretability: every prioritization decision traceable to explicit rules, with no black boxes

End-to-end ML pipeline analyzing vehicle sensor data (temperature, vibration, oil pressure, RPM) to predict failures before they occur.

• Multi-stage pipeline: data generation → ETL → EDA → ML → dashboard
• Random Forest classifier with imbalanced data handling (99.9% accuracy)
• Live Streamlit dashboard for real-time vehicle condition monitoring

Recommendation engine using matrix multiplication and Python multithreading to simulate scalable user-product scoring, with performance benchmarking between sequential and parallel execution.

• Implemented S = U × P matrix scoring with score normalization
• Compared sequential vs. multithreaded performance with visual output
• Demonstrated threading tradeoffs: overhead vs. scalability at scale