RosterAI: Workforce Optimization Engine

RosterAI. Most companies don't have a talent shortage. They have a routing problem. Matching engineers to projects requires understanding semantic overlap across documents, not just keyword matching. Project requirements live in slide decks and diagrams that most parsers skip. Managers assign greedily, locking the best candidate into the first project they evaluate. The result: top-heavy teams, understaffed initiatives, and people showing up with no idea why they were chosen. Roster AI treats this as a global constraint optimization problem. It ingests raw documents, extracts structured profiles, and scores every candidate against every project using a three-part hybrid model. A CP-SAT solver assigns everyone simultaneously. Resumes run through Docling. Slide decks get rasterized with PyMuPDF and processed by Llama 3.2 11B Vision Instruct via IBM watsonx.ai, extracting requirements from diagrams rather than typed text. A LangGraph agent enforces evidence anchoring: every recorded skill needs a verbatim source location or gets discarded. Scoring combines semantic similarity via multilingual-e5-large embeddings (30%), hard skill overlap with fuzzy matching (55%), and a rarity bonus for skills held by fewer than 10% of the workforce (15%). This stops the only PostgreSQL expert from landing on a team that just needs another React developer. The CP-SAT solver enforces team sizing, distributes senior engineers across projects, and treats pinned assignments as hard constraints. Every tradeoff surfaces as a Tension Alert. The React/TypeScript frontend explains each placement in one sentence, and drag-and-drop overrides immediately generate a consequence report.