AI systems
for real decisions.

Pre-departure delay prediction across U.S. flights.

A supervised learning system that scores scheduled flights using only pre-departure data. The pipeline compares model families, integrates point-in-time NOAA weather, calibrates probabilities and validates on chronological holdouts before producing a ranked operations queue.

Interpretable market regime detection with probabilistic time-series models.

A probabilistic time-series platform for five real assets using Gaussian Hidden Markov Models, engineered risk features and BIC-based model selection. Chronological validation and multi-seed stability checks assess robustness, while posterior probabilities, entropy and per-date feature evidence make each regime assignment inspectable.

Governed pricing optimisation from demand elasticity to executable price.

A decision-support system that estimates product-level price elasticity, simulates commercially feasible candidates and optimises under margin and volume constraints. Shared policy checks then determine whether a recommendation can execute automatically or must pause for persistent human approval, while FastAPI, Streamlit and the copilot converge on the same deterministic pricing and governance layers.

Evidence-backed account research and routing with agents, RAG and MCP.

An LLM-powered account intelligence workflow that combines first-party research with hybrid RAG, separates verified claims from hypotheses and routes accounts through structured decision rules. Its MCP-compatible agent interface preserves source provenance, evaluates retrieval on labelled queries and exports executive-ready briefs.

Probabilistic NLP for tracking narratives across the European AI Act debate.

An NLP pipeline that learns latent topics from attributable European AI Act sources using a custom collapsed-Gibbs LDA implementation. Source-held-out validation and cross-seed alignment test topic stability, while source-balanced profiles and Jensen-Shannon divergence compare narratives across actors in an interactive React atlas.

Constraint-aware rack placement for the HackUPC 2026 Mecalux challenge.

A team-built hackathon prototype that converts warehouse boundaries, obstacles, ceiling profiles and bay catalogues into feasible rack layouts under a fixed runtime budget. Its heuristic search combines multi-strategy shelf packing, access-gap reuse, residual-space filling and local repair, while the FastAPI interface provides 2D/3D inspection and evaluator-ready CSV export.