Mario H. Garrido Czacki

Fulbright–García Robles Scholar (2022).

• Leading the design of a JEPA-style energy-based foundation model for market panel data with structural inductive biases, supporting multi-horizon forecasting, intervention simulation, and probing of unobservable market drivers via synthetic supervision.
• Developed mechanisms to model latent, unobservable market drivers (e.g., price elasticity, cannibalization) via partial supervision from synthetic markets, allowing causal probing without direct observation in real data.
• Designed training objectives emphasizing structural inductive biases and compact representations, embedding key economic effects directly into the model and reducing reliance on brittle downstream estimation pipelines.
• Energy-based latent-variable architecture with JEPA-style training objective, trained on real market data, synthetic markets with known unobservable drivers, and distilled predictions from 4+ years of Kuona's production forecasting models. Modular adapters per client without full retraining — eliminating the 1–12 hour per-client cycle of prior approaches.

• Conceived and shipped ExpertAI end-to-end as sole architect — multi-agent backend (Django, Python), React frontend with WebSocket real-time communication, RabbitMQ/Kombu messaging system, cloud deployment (AWS EKS) — from concept to production in under a year.
• Multi-agent orchestration with compartmentalized context per subagent/tool, two-layer plan validation (formal PDDL gates + soft LLM-as-a-judge evaluators with scoped rule sets), and checkpoint-based resilience. Deployed and in use by global retailers and manufacturers.
• Designed a formal tool-planning framework where agents generate and execute multi-step plans verified via PDDL-style preconditions and effects, ensuring logical validity and dependency satisfaction before execution.
• Implemented deterministic plan reuse and semantic grounding mechanisms, enabling consistent execution, controlled plan adaptation, and resolution of ambiguous user inputs into unambiguous, executable specifications.
• Built Grafa: knowledge graph engine giving LLMs structured, relationship-aware memory over heterogeneous documents. Multi-tenant isolation with federated retrieval and hook-based extensibility at every pipeline stage.
• Built KQuery: NL-to-SQL engine across three database backends. Plans and verifies the full query hierarchy upfront — deciding intermediate schemas and SQL implementation details before execution — then runs all queries in parallel as soon as their dependencies are satisfied. Generation rules stored as data in the knowledge graph, scoped by platform, company, or user and updatable without code changes.
• Built KNER (Kuona Named Entity Recognition): multi-agent entity resolution pipeline that resolves free-text query segments against company-specific structured databases — products, geographies, promotions, brands, and business concepts — within a conversational, human-in-the-loop framework. Implemented as nested LangGraph state machines with four specialized LLM agents selected for cost/latency/capability tradeoffs.
• Designed UserAI, an autonomous evaluation harness that drives multi-turn agent sessions end-to-end — enabling CI-integrated regression detection of failure modes invisible in single-turn testing.
• Built an evaluation system that represents the agent as a directed graph of components, attributes failures to specific nodes via automated triage, ranks components by downstream harm for prioritized improvement, and provides deterministic replay for counterfactual experiments.
• Designed bypass auditing for agentic systems: when agents override validation gates, decisions are explicitly recorded — what was bypassed, why, and downstream consequences — making safety exceptions analyzable rather than invisible.

• Built vision-LLM quality monitoring systems and automated data acquisition pipelines over CommonCrawl, producing a ~700 GB high-quality dataset and training a custom SDXL ControlNet for conditional generation.

• Designed and deployed large-scale forecasting systems across thousands of market-correlated time series.
• Design, implementation, testing and deployment of a Machine Learning Pipeline that generates weekly demand forecasts for all of Heineken Mexico's product catalog in each of its distribution centers. Rewrote the entire system from scratch to improve results and performance for this process spanning 500+ servers, training approximately 3,500 recurrent neural networks weekly.
• Redesign and standardization of one of Kuona's main products, the Perfect Order demand forecasting system.
• Design and implementation of a new time series forecasting library featuring transparent multi-horizon predictions, automatic hyperparameter search, and native support for ensemble models.
• Design and implementation of a custom task orchestration system for Kuona's Machine Learning Pipelines with native support for heterogeneous distributed computing, arbitrary horizontal scalability, and easy management for remote tasks by abstracting the compute resources behind a UNIX-like process control interface.
• Regular meetings with clients to identify data requirements in order to guarantee consistent and precise forecasting. Evaluation of alternative datasources originating within clients' organizations.