Marcelo – Python, LLM, RAG

Production candidate-search platform for recruiters and hiring managers, combining retrieval, scoring, reranking, and explainable output to turn ambiguous hiring requests into auditable shortlist decisions. The system integrates semantic search, multi-criteria reranking, LLM-based candidate assessments, and chat-driven hiring workflows with a multi-agent AI runtime.

• Design and lead production AI search architecture combining Elasticsearch filters, semantic retrieval, and LLM-based assessments to improve shortlist quality while keeping results inspectable by recruiters.
• Build backend APIs, async workflows, and runtime infrastructure on Supabase/PostgreSQL with Redis-backed caching and queues to support high-throughput hiring operations.
• Implement chat-driven hiring workflows with planning, tool orchestration, memory-aware context injection, and AI SDK-compatible SSE streaming across backend and product surfaces.
• Design a PydanticAI-centric multi-agent runtime with specialist agent profiles, delegated subtasks, memory-aware shortlist re-evaluation, and feedback-triggered workflow adaptation.
• Build persistent LLM memory and session-aware context-injection infrastructure with confidence-scoped retrieval across sessions.
• Develop a trainable model layer for search with user-scoped persistence, quality tiers, SHAP/feature-importance-to-Elasticsearch-filter translation, and uncertainty-aware candidate scoring (Gaussian Process and conformal intervals).
• Strengthen runtime observability with request correlation, trace propagation, safe logging/redaction, and metrics/tracing patterns suited for debugging long-running agent workflows.
• Improve release safety through CI/CD hardening, test-driven refactors, and stronger observability across search, streaming, and agent-runtime components.

Internal assistant chatbots grounded on proprietary code and spreadsheet knowledge sources, with retrieval and prompt pipelines plus production guardrails for repeatable enterprise workflows.

• Built internal assistant chatbots grounded on proprietary code and spreadsheet knowledge sources with LangChain-based retrieval pipelines.
• Designed production guardrails and prompt optimization routines to ensure reliable, repeatable assistant responses aligned with organizational knowledge.
• Integrated LLaMA-based models with custom prompt engineering approaches to calibrate assistant tone and accuracy for internal use.

AI voice assistant with real-time transcription, TTS generation, and persona-aware RAG responses. Knowledge base built from a YouTube-to-vector-store pipeline with transcription, speaker diarization, and metadata-rich chunking.

• Led end-to-end development of an AI voice assistant combining real-time transcription, TTS generation, and persona-aware RAG response generation.
• Built YouTube-to-ChromaDB ingestion pipeline with AssemblyAI transcription, speaker diarization, and metadata-rich chunking for structured knowledge-base extraction.
• Implemented LangGraph agents with PostgreSQL checkpointing for persistent conversation state across sessions.
• Integrated ElevenLabs TTS, AssemblyAI STT, and Ultravox for real-time voice conversation handling, including audio event management (interruptions, pauses).
• Designed PCA-based scope analysis and 22-attribute concurrent persona extraction for personalized, personality-adjusted assistant responses.
• Built MinIO-synced ChromaDB vector store with Redis caching for production-ready retrieval infrastructure.

Computer vision platform for automated construction site monitoring using 360-degree photography. The system combined semantic segmentation of site elements with client-facing dashboards to track build speed and identify bottlenecks in real time.

• Led a team of data scientists building computer vision applications for automated construction site monitoring via 360-degree imagery.
• Developed a semantic segmentation solution to identify and classify construction site elements and track structural progress.
• Built client-facing dashboards enabling real-time visibility into construction speed and activity status.
• Designed the end-to-end pipeline from image ingestion through segmentation inference to dashboard visualization.

ML platform optimizing bidding strategies and campaign performance for digital marketing at scale, covering demand forecasting, anomaly detection, and A/B experimentation on large datasets backed by Snowflake and Amazon Aurora.

• Led redesign of bid-optimization and forecasting ML systems for digital marketing workflows.
• Built Python and SQL pipelines for production forecasting, anomaly detection, and decision support on Snowflake and Aurora-backed marketing data.
• Trained and tuned multi-level time-series models with Prophet-based forecasting and regression-driven adjustments, using cross-validation and quality gates (RMSE, MAE, MAPE, MASE).
• Implemented anomaly monitoring, alerting, and automated model-renewal logic for forecasting models in production.
• Supported bid-optimization and experimentation analyses, including A/B testing, confidence intervals, and statistical hypothesis testing.
• Partnered with business stakeholders to align model behavior with campaign KPIs and operational constraints.