Business decisions depend on more than numbers in tables. Reports, customer reviews, contracts, and support conversations contain explanations, concerns, and changes that structured datasets may miss. Turning this material into useful evidence requires more than asking AI for a summary: we need ways to organize the text, find relevant information, and check whether an interpretation is supported. This course explores how to build and evaluate those systems.
Build a pipeline. Defend its answer.
Learn AI engineering by building working pipelines with models, tools, and bounded agents. In Python, you will connect data processing, language models, and retrieval to investigate business questions—and evaluate how well your system answers them.
Designed for MSBA students, the course also welcomes students from other programs with instructor permission. It will especially interest students who want to explore finance, build practical AI workflows, or connect the two.
Preparation: Familiarity with generative AI tools is helpful. Students should be comfortable reading and modifying Python and applying basic data analytics. Code will be scaffolded, and the course introduces the model and agent concepts needed for the work. Contact me if you are unsure about your preparation.
Your building path
Prepare documents → Build a baseline → Add models and tools → Coordinate tasks → Evaluate the result.
Begin with a small document comparison. Extend it into a workflow that retrieves passages and produces structured, source-linked outputs. Then investigate routing and agent decisions, measuring whether each addition improves quality, cost, or latency.
In an agentic workflow, a model can choose a next step, such as retrieving another passage or using a tool, within limits you define. You will compare these systems with programmed approaches and learn to justify when the added complexity is useful.
What you will learn to build
- Document pipelines: clean and organize filings, reports, and other business text while preserving identifiers, dates, and links to the source.
- Classical NLP baselines: use tokenization, text representations, similarity measures, and classification to establish a comparison for model-based methods.
- Evidence-backed model outputs: develop prompts and structured extraction, retrieve relevant passages, validate outputs, and identify when an answer lacks support. This includes embeddings and retrieval-augmented generation, or RAG.
- Workflows and bounded agents: connect models to tools with defined inputs and outputs, manage context and state, and compare a fixed workflow with a model that chooses its next action. Inspect action logs and enforce retry limits and stopping conditions.
- Orchestration and routing: coordinate tasks and model calls, validate intermediate results, and decide when to retrieve more evidence, use another model, or ask a person.
- Model strategy and open weights: compare hosted and open-weight models on matched tasks, learn how inference and hosting choices affect a pipeline, and assess quality, cost, and latency tradeoffs.
- Evaluation and deployment decisions: design fair comparisons, inspect errors, and defend a recommendation to use, revise, or stop a system.
A shared business investigation
Corporate disclosures provide a running case: what changed in a company’s language, what does the evidence support, and when should a system retrieve more information or stop? Classroom work will use the Notre Dame filing collection to study preprocessing and larger document comparisons.
Can disclosure changes inform an investment strategy? One possible final project is to construct and evaluate a portfolio using changes in corporate disclosures. Compare it with a market benchmark and a text-based strategy inspired by Cohen, Malloy, and Nguyen’s Lazy Prices.
The scope depends on suitable filing and market data. Testing must account for when information became available, risk, and trading costs. This is an investigation, with no promised outperformance or full replication of the paper. A well-supported finding that a simpler method works better is valuable.
Projects can also investigate other business documents, such as reviews, contracts, or support records, with a focused question and usable data.
How you will work
Classes combine concepts, worked examples, and hands-on investigation. You will develop a pipeline in stages: establish a baseline, test changes, debug failures, and explain whether additional complexity improves the result.
Engineering evidence will accompany the code: small evaluation sets, model and prompt versions, tool-action logs, and measurements of failures, cost, and latency. You will use these records to explain why a design works, where it fails, and what should happen next.
The planned work includes four technical assignments, short evaluation and design memos, and a final project. Your final artifact will combine a working workflow, documented code, an evaluation report, and a demonstration. AI assistance is welcome; you remain responsible for understanding the code, prompts, and conclusions you submit.
BUS 244A, Gen AI Foundations for Business, is recommended preparation.
Try a public example
What changed in Apple’s AI risk disclosures? offers an existing example with original source passages and a Python comparison. It uses Python’s standard library and needs no model API access. The course builds described above are planned classroom work.
The broader Learn collection is open to everyone and will continue to grow alongside the course.
Interested in joining?
Contact me to discuss your interests and preparation. Students outside MSBA are encouraged to ask about permission to enroll. Industry colleagues interested in contributing a business question or sharing practical experience are also welcome to get in touch.