Marcos
Work

Projects

Selected AI/ML case studies covering the problem, my implementation, open-source foundations, and technical validation.

  • Automated extraction, validation, and human review for business documents.

    • Built an ingestion and extraction pipeline on the open-source Docling foundation
    • Added stateful human-review orchestration and workflow checkpointing with LangGraph
    • Designed business rule engines for 3-way matching, duplicate prevention, and mock ERP connectors
    Problem
    Invoices and purchase orders arrive in inconsistent formats and need extraction, review, validation, and delivery into a business workflow.
    My contribution
    Built the application layer around open-source Docling and LangGraph: ingestion, extraction workflows, human-review checkpoints, business-rule validation, duplicate checks, and mock ERP connectors.
    Validation / outcome
    Demonstrates a reviewable document-to-ERP workflow with checkpointed orchestration, rule-based validation, duplicate handling, and clear paths for human intervention.
    • Python
    • LangGraph
    • vLLM
    • Docling
    • Docker

    Need a document workflow adapted to your forms and business rules? Discuss your project.

  • Secure search across private company documents.

    • Implemented a FastAPI Copilot API Gateway to handle authentication, ACL resolution, and query routing
    • Integrated the open-source RAGFlow foundation 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
    Private company knowledge requires retrieval that respects user and group permissions while preserving traceability across queries.
    My contribution
    Built the gateway, permission-resolution, query-routing, and audit layers around RAGFlow, an open-source retrieval foundation, with a PostgreSQL ACL model.
    Validation / outcome
    Demonstrates permission-aware retrieval, hybrid and graph-based search paths, and auditable request handling for private knowledge systems.
    • Next.js
    • FastAPI
    • RAGFlow
    • PostgreSQL
    • GraphRAG

    Need secure retrieval added to an existing product or document stack? Discuss your project.

  • Optimized serving, benchmarking, monitoring, and security testing for LLM deployments.

    • Developed a model configuration registry and deployment API around the open-source vLLM serving engine
    • Automated a quantization experiment matrix (BF16, AWQ, FP8) and load testing benchmark suite
    • Integrated the open-source Garak scanner with custom probes for prompt extraction and canary leakage
    • Created a CI security gate that blocks deployments automatically on performance or security degradation
    Problem
    LLM prototypes need repeatable deployment, performance comparison, monitoring, and security checks before they can be operated reliably.
    My contribution
    Built deployment orchestration, configuration management, quantization experiments, load testing, monitoring, and release gates around the open-source vLLM and Garak foundations.
    Validation / outcome
    Demonstrates a repeatable process for comparing serving configurations, observing performance, running security probes, and blocking releases that fail defined checks.
    • vLLM
    • Garak
    • Prometheus
    • Grafana
    • FastAPI
    • Docker

    Have an LLM prototype that needs benchmarking, hardening, or deployment? Discuss your project.

  • Real-time personalized ranking with cold-start handling.

    • 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
    Personalized ranking needs fresh behavioral data, a cold-start path, reproducible training, and visibility into model drift.
    My contribution
    Built the streaming feature flow, two-stage retrieval and ranking pipeline, cold-start handling, offline evaluation, inference API, and monitoring integration.
    Validation / outcome
    Demonstrates real-time feature updates, two-stage recommendation, offline evaluation, online inference, and drift monitoring in one system design.
    • FastAPI
    • Kafka
    • Redis
    • PostgreSQL
    • LightGBM
    • LightGCN

    Need ranking or recommendations added to an application with live data? Discuss your project.

  • Detection and tracking across camera streams.

    • Built a multi-camera stream manager and frame batcher synchronized across RTSP streams
    • Integrated the open-source Ultralytics YOLO detector and ByteTrack/BoT-SORT trackers
    • 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
    Multi-camera operations need synchronized detection and tracking so events can be reviewed without searching entire video streams manually.
    My contribution
    Built stream synchronization, event rules, searchable clip generation, and system integration around open-source YOLO detection and ByteTrack/BoT-SORT tracking foundations.
    Validation / outcome
    Demonstrates synchronized multi-camera processing, consistent object tracking, configurable zone rules, and event-linked video retrieval.
    • YOLO
    • ByteTrack
    • Python
    • PostgreSQL
    • Redis
    • OpenCV

    Need an image or video pipeline integrated with your existing operations? Discuss your project.