A campus map shows where a building is, not which door you can get through. AccessiGo classifies an entrance photo as accessible, uncertain, or not a usable entrance, and saves the result to the user's account for the next trip. The uncertain band is deliberate: ambiguous doors go to human review rather than a guess.

The model is a Keras CNN trained on 256×256 entrance images and evaluated on precision and recall, not accuracy alone. TensorFlow was too large for the target serverless environment, so the trained model exports to ONNX and serves through ONNX Runtime behind a parity check that rejects the artifact if predictions drift past 0.00001. Measured drift came in at 5.96e-08. Migrating the backend from Flask to FastAPI preserved the existing upload contract, so the already-released React Native client kept working without a coordinated deploy.

Financial research punishes confident guesses, so this is built to refuse rather than guess. Ask a question in natural language across 25 official 10-K filings from Apple, Microsoft, NVIDIA, Amazon, and Alphabet, and every claim returns tied to a passage you can open and read.

Citations work as a validation contract rather than decoration. The model can only cite evidence registered for that specific question. The validator maps each identifier back to its canonical passage, checks the quoted support and any financial values, allows one correction attempt, then returns a no-evidence response if the answer still can't be supported.

Retrieval pairs pgvector semantic search with PostgreSQL full-text search through Reciprocal Rank Fusion, so paraphrased questions and exact accounting terms both find their target. SEC tables were the hard part, since a generic HTML converter makes a filing readable while destroying the link between a number and its row label, unit, and period. A deterministic lxml extractor keeps that structure intact and indexes every financial row as its own retrieval chunk.

A knight moves through shifting magic terrain, traps, and room puzzles. Written without an engine, so the game loop, collision handling, and sprite state machines are all hand-rolled.

Procedural mazes and puzzles with no framework and no build step. Rendering, input handling, and level state are written directly against the DOM and canvas.