Flagship
AI Decision Copilot
A tabular analytics agent — profiling, investigation, explainable prediction, scenarios — not a chat-over-CSV demo.
- Problem
- A CSV lands. Without a method you get a plausible summary. The need is a loop of investigation → explanation → simulation, with tool outputs before prose.
- Why it matters
- This is the Lab’s Data → Agents bridge. The repo README positions it as an analytics product, not a chatbot demo.
Stack: Python · FastAPI · pandas
Architecture
Documented flow: upload → diagnose → investigate → explain → simulate → recommend → export.
FastAPI serves the UI. Ingestion, profiling, investigation, root-cause, ML engine, scenarios, actions, copilot and HTML export are separate modules, not one prompt.
Data
Uploaded CSV or the sample set. Derived features in the profiler. No client-data claims.
Approach
- Profile before narrating.
- Rank investigation paths.
- Train an explainable model when a target exists (the demo uses
revenue). - Compare scenarios before recommending.
- Export an HTML analysis note.
Results
The profiling lab publishes an A/B/C run (profile-llm.v1): sample vs tool vs extracted claims. Agent scores outside that protocol stay off this page.
Limitations
The README is explicit: portfolio / short demo / business storytelling. Not a client deployment. No full-workflow task-completion score.
Learnings
A credible analyst agent needs tools and artifacts (profile, scenarios, report) before prose. The sample-vs-tool gap is now numbered in the Lab.
Next
Overwrite trace C with a live call (python -m labeval --run profile-llm --live) once a key is available.