Marcos
Work

Projects

Selected products, open source libraries, and AI work — what the problem was, my role, and the outcome.

  • Automated pipeline converting unstructured documents into actionable ERP data.

    • Built a document ingestion pipeline with baseline and fine-tuned extraction models using Docling
    • Developed a stateful orchestration layer via LangGraph for human-in-the-loop review and workflow checkpointing
    • Designed business rule engines for 3-way matching, duplicate prevention, and mock ERP connectors
    Problem
    Processing unstructured documents like invoices and purchase orders required manual labor, leading to errors, financial losses, inventory discrepancies, and delayed vendor payments.
    Role
    Designed and built the end-to-end pipeline, covering document ingestion, model extraction, human review orchestration, business-rule validation, and ERP integration.
    Result
    Reached 98.2% total extraction accuracy and established a reliable automated document processing system capable of blocking duplicate attempts with a 99.9% successful completion rate.
    • Python
    • LangGraph
    • vLLM
    • Docling
    • Docker

  • Private enterprise knowledge copilot with RAG, GraphRAG, permissions, and evaluation.

    • Implemented a FastAPI Copilot API Gateway to handle authentication, ACL resolution, and query routing
    • Integrated RAGFlow for document parsing, embeddings, vector search, and GraphRAG capabilities
    • Developed a PostgreSQL-based ACL system to manage users, groups, permissions, and audit events securely
    Problem
    Organizations needed a secure enterprise knowledge assistant with robust access control, preventing unauthorized data access while providing accurate, graph-based answers.
    Role
    Designed and built the end-to-end copilot, including authentication, permission resolution, query routing, retrieval orchestration, and secure audit tracking.
    Result
    Created a secure enterprise copilot system capable of answering queries based on complex permissions while supporting hybrid and graph-based retrieval modes.
    • Next.js
    • FastAPI
    • RAGFlow
    • PostgreSQL
    • GraphRAG

  • High-throughput model serving platform with integrated security probing and automated quantization.

    • Developed a model configuration registry and deployment API to orchestrate vLLM instances
    • Automated a quantization experiment matrix (BF16, AWQ, FP8) and load testing benchmark suite
    • Integrated Garak for custom security probes targeting system prompt extraction and canary leakage
    • Created a CI security gate that blocks deployments automatically on performance or security degradation
    Problem
    Deploying LLMs in production requires optimizing for throughput, ensuring security against prompt extraction and leakage, and preventing performance degradation across configurations.
    Role
    Designed and built the end-to-end serving platform, including deployment orchestration, quantization experiments, load testing, security probes, and automated release gates.
    Result
    Delivered an optimized LLM deployment pipeline that securely proxies requests while ensuring high inference throughput and passing canary leakage tests with zero failures.
    • vLLM
    • Garak
    • Prometheus
    • Grafana
    • FastAPI
    • Docker

  • Two-stage production recommender using candidate generation, ranking, and streaming feature updates.

    • Built an event-driven recommendation API gateway with streaming feature updates via Redpanda/Kafka and Redis
    • Implemented a two-stage inference pipeline using LightGCN for candidate generation and LightGBM for ranking
    • Integrated Evidently to track offline evaluation, production drift, and recommendation metrics
    • Developed a full MLOps training pipeline, cold-start resolution system, and monitoring dashboards
    Problem
    An event-driven recommendation system was needed to process real-time streaming data while connecting cold-start handling, offline training, and production model drift tracking.
    Role
    Designed and built the end-to-end recommendation platform, spanning streaming features, candidate generation, ranking, cold-start handling, evaluation, and production monitoring.
    Result
    Successfully launched a robust recommendation platform with end-to-end MLOps pipelines supporting both offline evaluation and real-time inference.
    • FastAPI
    • Kafka
    • Redis
    • PostgreSQL
    • LightGBM
    • LightGCN

  • Real-time multi-camera tracking and detection system for event generation and zone monitoring.

    • Built a multi-camera stream manager and frame batcher synchronized across RTSP streams
    • Integrated Ultralytics YOLO for object detection and ByteTrack/BoT-SORT for consistent tracking
    • Developed an event rules engine using Supervision to evaluate zone entries and physical constraints such as forklift proximity
    • Set up automated searchable clip generation, storing video event slices to MinIO (S3) for a monitoring dashboard
    Problem
    Warehouses require robust multi-camera surveillance to automatically detect events, enforce restricted zones, and monitor proximity safety without intensive manual review.
    Role
    Designed and built the end-to-end video intelligence pipeline, covering synchronized streams, object detection, multi-camera tracking, event rules, and searchable clip generation.
    Result
    Delivered a reliable, real-time video analytics pipeline capable of multi-camera tracking and automated alert generation for physical warehouse safety zones.
    • YOLO
    • ByteTrack
    • Python
    • PostgreSQL
    • Redis
    • OpenCV