Projects

Capstone briefs

End-state projects for each track, plus an optional multimodal convergence.

Path A

Fine-tune or LoRA-tune + model card

Train or adapt an open model; ship an eval story, not just a notebook.

Curated public or licensed dataset with documented splits.
Baseline vs adapted model with automatic metrics + small human rubric.
Training config, pinned deps, checkpoint pointer, 1–2 page model card.
Optional: inference note (latency/memory) before/after a lightweight quant setup.

Path B

Internal copilot — RAG + LangGraph + tool

Ground answers on docs, orchestrate explicitly, measure cost.

Chunked corpus with metadata (path, ACL group, freshness).
Vector + hybrid keyword retrieval; citations required in answers.
LangGraph (or equivalent explicit graph) with retrieve → draft → verify → optional tool.
One safe, schema-defined tool with idempotency patterns.
Written $/1k session estimate at your model tiers; trace IDs across steps.

Shared

Optional multimodal agent

After either capstone.

Accept screenshot/PDF excerpt; choose VLM vs OCR deliberately; keep citations when using text corpus; document five failure modes and how the system degrades.

Syllabus anchors: Path A capstone slot · Path B capstone slot