Build AI systems.
Investigate how they work.
A public learning collection on AI workflows, agents, and open-weight models. Start with prompting, evidence checks, and saved local-model examples you can inspect. Retrieval, agents, and routing are next as the collection grows.
New detail.
New risk?
Apple revised its discussion of third-party intellectual property between its 2024 and 2025 annual reports. What changed in the disclosure—and what would you need before drawing a business conclusion?
Selected phrases from Item 1A, Risk Factors. Read the full paragraphs in the example.
Runs with Python’s standard library; no model API needed.
“can generally be obtained on commercially reasonable terms”
The same paragraph also says availability is not assured.
Compare the 2025 passage
“not always able to obtain all necessary licenses”
The paragraph also adds specific AI risks involving copyrighted training materials and their potential reproduction in outputs. Does that establish a newly occurring problem? Check the full passages before deciding.
Same AI technology.
Different business risks?
Apple. Microsoft. Alphabet. Meta. All four discuss AI—but does the same language describe the same business problem?
Build a comparison from original filings. Find a shared mechanism, defend a difference, and check whether either is new.
Compare the companies- Read across companies.Look beyond the shared keywords.
- Check both years.A distinctive risk may already be familiar.
- Defend the connection.Attach evidence to each business claim.
12 source selections from 4 companies.
Python 3.9+, with optional AI assistance.
Give a local AI model
a useful job.
Qwen turned two Apple disclosures into claims, quotations and a comparison. What would you keep—and what would you change?
Explore a real saved result and the choices behind running a model yourself: control, customization, data handling and the resources it needs.
Explore the Qwen example- The model organizes.Claims with source passages you can inspect.
- Python checks.Structure, exact quotations and year coverage.
- You judge.Keep, revise or reject an interpretation.
Saved local Qwen run and offline Python checker.
No model installation needed.
From one experiment to a working system.
Can a system identify meaningful changes between company reports, retrieve supporting passages, and recognize when the evidence is insufficient?
Start with the Apple comparison, try the peer exercise, then inspect a model-assisted extraction and its evidence checks. I plan to extend this work through full-document retrieval, bounded agents and model routing. The same engineering questions apply across document domains. Three strands will guide the collection:
- 01
Who chooses the next step?
Workflows and agents: First workflow available
The extraction example follows a predefined sequence with a separate support review. Next, investigate when a model should choose an action, what tools it can use, and when it should stop.
- 02
When should a system ask for help?
Orchestration and routing: Planned
Explore how to divide a task, check intermediate results, and decide whether another model or a person is needed.
- 03
What changes when you run the model?
Open-weight models: First saved example available
Start with the local Qwen example, then inspect a small adapter-training result: did changing weights help more than clearer instructions? Explore deployment choices, evidence checks and the limits of each result.
Build it. Test it. Explain your choice.
BUS 243A is planned for Spring 2027: a classroom path through AI engineering with unstructured business data. Interested in finance, practical AI, or both? Students outside MSBA can contact me about permission.
Explore the course →