Projects

Products I’ve built end to end — not a checklist of subsystems.

AI Planning & Procurement Platform · Valerdat

The planning and procurement platform I led and designed hands-on, and developed together with the Valerdat team — demand planning, purchasing, and replenishment: ERP supply-chain data in, forecasts and purchase proposals out. On top, agents don’t just chat: they automate work — querying demand, inspecting proposals, and taking actions in the platform.

1 The first system was slow and unstable even for a small client. Tight coupling between the early modules blocked rollouts and further development. The rewrite separated those modules.
Re-architected the platform end to end — Spark pipelines and demand forecasting for planning and procurement.
Owned the full stack: distributed ETL (Databricks/Spark), a multi-tenant Django backend, and AWS infrastructure, observability, and security.
2 Built to cut implementation friction — a hard limit on business scalability.
Delivered multi-ERP connectors — Sage 200, Microsoft Dynamics 365 Business Central — via API/SFTP, listed on the Sage Marketplace.

Architecture

How the platform goes from ERP data to purchase proposals: the system analyzes forecasting models as the base for planning, a purchase optimizer turns that into buy proposals, and the platform is the base system each client builds on. Agents on top converse and automate, not only answer questions. The whole path is wrapped in each client’s business logic: their rules and constraints shape ingest, models, the purchase optimizer, the platform, and what agents are allowed to do.

From ERP data to purchase proposals

Wrapped in each client’s business logic

Select a stage — every step runs inside the client’s business logic.

Business logic. Every stage below is tenant-specific: each client’s rules and constraints wrap ingest, models, the purchase optimizer, the platform, and agent actions.
Ingest. Connectors pull from SAP, Dynamics 365 Business Central, and Sage into S3. Databricks jobs validate the data, apply each client’s rules, and prepare features for forecasting.
Forecast. The system compares forecasting models and keeps the best fit per SKU and warehouse. That signal is the base for planning.
Purchase optimizer. Turns the forecast into purchase proposals under each customer’s business constraints — the technical optimization step.
Platform. Results land in PostgreSQL and surface in the Angular app. It is also our base system: clients build their own system on top of it.
Agents. A layer on top of the whole stack: natural-language conversation plus automations — query demand, inspect proposals, and take actions in the platform, always within that client’s business logic.

Strategic Planning Platform · Inditex

Internal planning product for Inditex’s Security Department — Django services and a React UI on AWS, with APIs and data paths shaped for how that team actually plans and operates.

Open-source and side projects — PyPulse, cefpythonX, real-tars, and more — live under Research.