Mario H. Garrido Czacki

Fulbright–García Robles Scholar (2022).

• Defined the system decomposition and architectural structure for Kuona's forecasting platform, redesigning it from client-specific pipelines into generalized, reusable infrastructure. This architecture has been followed by the data science team for approximately two years and serves as the basis for the company's foundation model development.
• Decoupled system components into composable ETL pipelines, eliminating hardcoded client-specific workflows. Promoted anomaly detection and outlier tracking to first-class persistent training signals across datasets. Removed architectural bottlenecks and standardized infrastructure across forecasting systems.
• Designed generalized inference infrastructure supporting immediate forecasting without per-client retraining, with modular per-client adapters eliminating 1–12 hour retraining cycles. Reduced onboarding time for new client data from days to hours. Real-time forecasting capabilities enabled by this architecture are in active development.
• Leading the design of a JEPA-style energy-based foundation model on this platform — learning compact representations of unobservable market dynamics via synthetic supervision, supporting multi-horizon forecasting, intervention simulation, and causal probing.

• Defined the ExpertAI system architecture, which is now the required architectural foundation through which all agent initiatives at Kuona operate. All agent-based systems and initiatives follow the execution model, infrastructure, and architectural structure defined in this system.
• Other teams build on top of this architecture — agent initiatives across the organization depend on the execution, orchestration, context management, plan validation, evaluation, and resilience patterns established in ExpertAI.
• Concept to production in under a year. Multi-agent analytical platform — 90% task success rate, reduces cross-silo queries from hours to minutes across 5 database backends. Full-stack: Python/FastAPI + Python/TypeScript/Django backend, React frontend, WebSocket real-time communication, RabbitMQ/Kombu messaging, AWS EKS deployment.
• Designed the evaluation and observability infrastructure for agent systems: agents represented as directed graphs with automated failure attribution to specific nodes, component ranking by downstream harm, and deterministic replay for counterfactual experiments.
• Designed formal tool-planning framework where agents generate multi-step plans verified via PDDL-style preconditions and effects, with bypass auditing that explicitly records safety exception decisions for post-hoc analysis.
• Implemented deterministic plan reuse and semantic grounding mechanisms, enabling consistent execution, controlled plan adaptation, and resolution of ambiguous inputs into executable specifications.
• Built Grafa: knowledge graph engine providing LLMs structured memory over heterogeneous documents. Sub-second federated retrieval, ~95% recall. Multi-tenant isolation with hook-based extensibility. Deployed as shared infrastructure across agent products.
• Built KQuery: NL-to-SQL engine across five database backends, ~95% SQL accuracy. Plans and verifies the full query hierarchy upfront, then runs all queries in parallel as dependencies resolve. Generation rules stored as data in the knowledge graph, scoped by platform, company, or user.
• Built KNER: multi-agent entity resolution pipeline resolving free-text queries against all company entity systems — products, geographies, promotions, brands, ontological concepts — in 10–20 seconds including disambiguation. Implemented as nested LangGraph state machines with four specialized agents.
• Designed UserAI, an autonomous evaluation harness driving multi-turn agent sessions end-to-end for CI-integrated regression detection of failure modes invisible in single-turn testing.

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

• Replaced Kuona's fragmented per-client training-on-demand systems with a unified, generalized distributed ML training platform — deprecating legacy systems entirely and deploying a single platform spanning 500+ servers, supporting thousands of model training jobs weekly, with orchestration infrastructure, training pipelines, and monitoring systems.
• This structural architectural replacement reduced onboarding time for new configurations and clients from weeks to days, improved system scalability and reliability, and improved forecasting accuracy through standardized training infrastructure.
• Designed and implemented a custom task orchestration system with native support for heterogeneous distributed computing, arbitrary horizontal scalability, and a UNIX-like process control interface for remote task management.
• Led a 3-person team as lead architect in the redesign of Kuona's Perfect Order demand forecasting system. Introduced pre-modeling data governance checks that caught data errors upstream, saving hundreds of GPU hours in wasted training.
• Established engineering infrastructure and practices adopted across engineering and data science teams: eliminated development-to-production drift, implemented ephemeral environments and reproducible infrastructure, and standardized workflows and deployment environments across production ML systems.
• Designed and implemented a time series forecasting library with transparent multi-horizon predictions, automatic hyperparameter search, and native ensemble model support.