Course demonstration · Capstone 2 · Agentic AI Workshop 2026

Product

Platform

web

Stack

Python back end (the existing app/analyst package: tools, agent loop, replay, scorecard, briefing) served locally by FastAPI; a hand-built HTML/CSS/JS front end with no framework. Runs entirely on the presenter's machine. Chosen by the user over restyling Streamlit, for full design control.

Users

Judges and the "leadership team" at the Agentic AI Workshop 2026 capstone presentations: they watch a 6-minute live demo projected in a lit workshop room, read from a few metres away, then ask questions live (including one that leads to a planted anomaly). They also see a recorded video.
The presenting team (mixed, many not Python users) drive the demo: click prepared cases, type live questions, fall back to replays if the AI or wifi fails.
In the story the product tells: managers who ask simple questions about the numbers and today wait days for an analyst.

Product Purpose

A Business Performance Analyst Agent: a manager asks a plain-English question about sales, and the agent plans the analysis, writes and runs read-only SQL or pandas, corrects its own errors, draws a chart, explains what drove the result, flags anomalies, and writes a one-page weekly leadership briefing. Success at the presentation means judges see a working demo they can trust, and score it on business value, agent design, working demo, trust & safety, and pitch.

Positioning

Every number is traceable: the exact queries that produced an answer are shown from the tool log, not from the AI's own words, against a read-only database. The agent says when the data cannot answer and never presents a forecast or guess as a fact.

Operating Context

Live presentation: 13:45-15:30 sprint, 15:45-16:30 presentations; each a 6-minute demo plus 2-minute Q&A.
Projector in a lit room; also a recorded, narrated video produced by app/scripts/record_video.py.
Modes: Live AI (Groq free tier, rate-limited, with automatic fallback to OpenAI gpt-5.4-mini when the daily allowance runs out), Replay of a recorded live run, Dry run (no AI, scripted, for testing only). Replay and dry run are always labelled as such on screen.

The same screens serve three audiences, and the mode is always labelled on screen:

flowchart LR
  accTitle: Who sees what on the day
  T["Presenting team"] --> M{"Mode"}
  M -- "AI and wifi working" --> L["Live AI<br/>Groq, then OpenAI"]
  M -- "AI or wifi fails" --> R["Replay of a<br/>recorded live run"]
  M -- "testing only" --> D["Dry run<br/>labelled in red"]
  L --> J["Judges, on the projector<br/>(6 min demo, 2 min Q&A)"]
  R --> J
  R --> V["Narrated video"] --> J
  classDef red stroke:#c62828,stroke-width:2px,color:#b71c1c
  class D red

Capabilities and Constraints

Data: Microsoft AdventureWorks (MIT), 30 May 2022 to 29 Jun 2025, 31,465 orders, US dollars, 10 sales territories, Online and Reseller channels, 4 categories. Local read-only SQLite.
Tools: get_schema, run_sql, run_python, make_chart, detect_anomalies, explain_change.
Margin = LineTotal - (OrderQty x StandardCost).
Screens: overview, what changed (before/after vs the starting notebook), demo cases, tools without AI, weekly briefing, scorecard.
The video recorder drives the UI by URL parameters and waits for a scene-complete signal (scripts/record_video.py):

page.goto(f"http://localhost:{PORT}/?{params}&mode={args.mode}&presenter=1&autorun=1&delay={args.delay}")
page.wait_for_selector("body[data-scene=done]", timeout=600_000)

Brand Commitments

Name: "Business Performance Analyst Agent" for AdventureWorks. No team branding. No workshop or organiser (sensiwise.ai) branding, and nothing that impersonates Microsoft.

Evidence on Hand

Real query results from AdventureWorks; recorded live runs in app/recordings/; the scorecard. No customer testimonials, benchmarks or business results exist; none may be invented.

Product Principles

Show the working: the exact queries sit under every answer ("How I got this", in Developer view).
Honest about limits: "the data does not include…" is a feature, shown with the same weight as an answer.
Never blur modes: live, replay and dry run are always distinguishable.
Readable from the back of the room.

Accessibility & Inclusion

Projected display: large type, high contrast that survives a washed-out projector, no meaning carried by colour alone.

A learning project, not a product. AdventureWorks is Microsoft's public sample data about a made-up bicycle company. Made by Victor Saly, Akashdeep Nijjar, Manuel Verduzco Valenzuela, Mazen Ahmed, Buddhika Gamage, Nathan Fryatt, Michael Kampouridis and Malak Sheat (the team). Source of this page: docs/product.md.