Syllabus

Course description

Applied AI studies the engineering work between a model's capabilities and a system that people can use. The course moves from model behavior and prompting through retrieval, evaluation, tool protocols, agent workflows, efficient deployment, and safety. A capstone brings these ideas together in a working system with measured results.

This is a public self-study edition. The course has also been taught at the University of Connecticut in Spring 2025 and Spring 2026. Public access provides the course materials; university registration, academic credit, instructor assessment, and certificates are not part of this edition.

Learning outcomes

By working through the materials and practice activities, you should be able to:

1. Define a useful AI task and compare model choices using an explicit evaluation plan.
2. Build prompts and tool interfaces with validated inputs, structured outputs, and failure handling.
3. Design retrieval and grounding pipelines and distinguish retrieval failures from generation failures.
4. Evaluate AI behavior with task metrics, calibrated model judges, and human review.
5. Explain MCP's protocol responsibilities and implement a tool integration with appropriate authorization.
6. Choose an agent workflow, manage persistent state, and define where human approval is required.
7. Compare model size, fine-tuning, routing, latency, and cost using measurements from the actual task.
8. Test prompt injection and other failure modes, apply layered controls, and communicate remaining limits.

Preparation

The materials assume that you can read and write Python, use Git and a command line, call an HTTP API, and work with JSON. Basic familiarity with machine learning, probability, and vector similarity will help. The server examples also include TypeScript. Earlier lectures introduce the context needed for later topics.

How to study

Follow the 15-week schedule in order, or use the weeks as modules at your own pace. Each week offers a reading view, the original slide deck, and suggested practice. Read the lecture, inspect its cited sources, complete an experiment, and keep a short record of what worked and what failed.

The suggested practice on the schedule is a self-study guide. Existing lecture exercises are retained in context. Begin thinking about a capstone during the retrieval and evaluation modules; the suggested milestones start in Week 10.

Materials and computing

All published lecture materials are available on this site without an account. No textbook is required. Papers and provider documentation are linked from the lecture notes and resources page.

Some experiments need a model API or local inference software; hosted APIs and rented GPUs may incur charges. Start with a small dataset, set spending limits, and choose hardware appropriate to the model's full memory requirements. Record the exact model, API, package versions, and experiment date. No starter projects, datasets, videos, or hosted compute are promised beyond the files actually published here.

Capstone and assessment

The capstone is a working application with a reproducible repository, a demonstration, an architecture writeup, evaluation results, and a failure analysis. The capstone guide publishes the existing criteria and suggested milestones. Use its rubric for self-assessment or peer feedback. The percentages describe that project's dimensions, rather than an overall university course grade.

Responsible use and attribution

Document AI assistance, cite the work you build on, and retain the provenance of datasets and synthetic examples. Use data you are permitted to process. Keep credentials and personal information out of repositories and demonstrations. Exercise destructive tools only against disposable resources you control.

Version notes

The public site was revised on October 20, 2026. The lecture collection includes examples from the taught editions; model names, API signatures, pricing, and protocol versions can change. Update notes identify reviewed corrections and known areas requiring further revision. Consult those notes before running the MCP exercises or choosing deployment hardware.

Accessibility and reuse

Every lecture has an HTML reading view alongside the slide viewer. Course documents and notes support printing or saving as PDF from your browser. If an embedded diagram is unavailable, its source remains available in the downloadable lecture Markdown. Report an issue through the instructor's contact page.

Course materials are copyright Sawan Ruparel. Lecture content may not be redistributed without permission. These materials retain their existing reuse terms.

Representative study exercise

This exercise is newly written for the public edition, drawing on Week 7: Evaluation, particularly its discussions of faithfulness, evaluation datasets, and human review. It is an example of how to study the material, rather than a claim about an assignment used in either university offering.

Question: Can a model answer a policy question faithfully when the supplied source does not contain the answer?

Use this small, fictional policy passage: “Workshop registration closes on Friday. Registered participants receive the preparation guide by email on Monday.” Ask: “Can I receive a refund if I cancel on Saturday?”

Compare these illustrative responses; neither is a measured model result:

• Response A: “Yes. Cancellations within 24 hours qualify for a refund.”
• Response B: “The passage does not state a refund policy. Contact the workshop organizer before assuming a refund is available.”

Response A invents a rule absent from the source. Response B identifies the missing evidence and offers a next step. Both discuss refunds, so relevance alone would miss the failure. Faithfulness distinguishes them.

Try the following small experiment:

1. Create ten questions from a short document you are permitted to use: five answerable from the text and five requiring information it omits. Write reference answers or explicit “insufficient evidence” labels before running a model.
2. Compare a baseline prompt with a prompt that asks the model to cite supporting passages and abstain when evidence is missing. Keep the document, model version, and generation settings fixed.
3. Review each answer manually for supported claims and correct handling of missing evidence. Record the case, prompt version, answer, judgment, latency, and API cost where available. Report counts with their denominator.
4. Inspect disagreements and tradeoffs. Did abstention reduce invented claims while also refusing answerable questions? Explain one remaining failure and what you would test next.

Evidence to keep: the ten cases, both prompt versions, raw outputs, a case-level results table, and a short analysis. Ten cases support an initial diagnostic exercise; they do not establish general reliability. A model judge can assist a later pass, but check its judgments against your own labels.

Adapting the course for an institution

The course was taught at the University of Connecticut in Spring 2025 and Spring 2026. This public syllabus describes the independent-study edition. It does not announce another institutional offering or reproduce an official university syllabus.

The published sequence provides a starting point for a graduate course: preparation in Python, Git, HTTP APIs, and basic machine learning; fifteen ordered lectures; suggested practice; and a capstone with explicit evaluation and reproducibility criteria. An adopting instructor should map the learning outcomes to local program goals and available teaching time, then choose which exercises and capstone milestones to assess.

Before an institutional offering, establish the course identifier, calendar, credit hours, contact and office hours, attendance and submission policies, overall grade weights, accommodations, academic integrity and AI-use rules, and access to affordable compute. Those arrangements are institution-specific and have not been established by this public edition. Review dated technical examples and choose compatible dependency versions before assigning labs.

The published capstone percentages apply to that project's dimensions; they do not prescribe an overall course grade. Course materials retain their existing copyright terms. Contact Sawan Ruparel to discuss teaching, adaptation, or reuse permission.