Applied LLMs — origins, evidence and applications

Who this is for

You can write basic Python and have used a chat model, but want to explain tokens, attention, retrieval mistakes and deployment responsibilities rather than collect prompt folklore. This is a reading and practice map, not a claim that reading eight pages deploys a reliable service. The full route retains foundations, mechanics, retrieval, prompting, audio, clinical NLP, statistical grounding and a cloud platform branch.

Start with concepts; calculate one example; then design a failure test. The foundations page needs no calculus. Attention uses vectors and weighted sums, with shapes explained in its own lesson. The optional classical-ML route supplies splits, losses and evaluation vocabulary whenever you need it. No paid API, cloud account or accelerator is required merely to study this path.

The path

Part 1 · Foundations

Where LLMs came from, and what they actually are traces count models, neural representations, sequence models, attention, scaling and post-training. It distinguishes a model from the software around it and next-token prediction from other objectives.

Exit test: explain why changing a prompt, saving product memory and updating weights are three different operations. Identify one fluent claim whose source you would check. This is the prerequisite vocabulary for the rest, not a demand to finish every optional video.

Part 2 · Mechanics

How transformers actually attend introduces attention as a weighted lookup. Then use self-attention for the full calculation and multi-head attention for shape/mask controls.

Exit test: calculate a two-value weighted sum, explain the causal mask, and separate quadratic pairwise arithmetic from whether a kernel materializes its score matrix. A larger context window is neither free nor proof of accurate retrieval across that window. Read part1 first; use the math prerequisites in the linked lessons when needed.

Part 3 · Retrieval

RAG architecture, evidence and evaluation covers document parsing, chunking, embedding contracts, vector indexes, permission filtering, hybrid ranking, reranking, generation and separate retrieval/support metrics. Its exact lexical fixture supplies a visible baseline; its local integration recipe states which external components remain unexecuted.

Exit test: diagnose a related-but-wrong chunk, a missing exception at a chunk boundary, no evidence, and a correct citation ID attached to an unsupported answer. Explain why authorization happens before the generator sees text. Part1 is enough to begin; part2 helps you reason about encoder/context limits.

Part 4 · Prompting

A deep-study prompt with examples and evidence is a practical way to request mechanism, examples and checks. Pair it with zero-shot, few-shot, reasoning and ReAct prompting when comparing strategies.

Exit test: write a task specification containing the desired output, necessary context, constraints, an example and a verification criterion. Compare outputs on fixed cases. A prompt that asks for depth does not guarantee it, and “think step by step” is not a universal improvement or a security boundary.

Part 5 · Applied audio

Automating audio generation follows text through Azure Speech synthesis, file handling, Blob Storage and a web player. The lesson distinguishes synthesis success from upload/read authorization and identifies unexecuted service boundaries. It is a case study/reference, not a measured production throughput report.

Exit test: if synthesis succeeds and upload fails, say which artifact must remain available and how a retry avoids unnecessary regeneration. Define pronunciation/pacing review, format metadata, access controls and a measured cost plan. Read parts1–3 first; this branch also benefits from basic HTTP and file I/O.

Part 6 · Clinical NLP

Reading MEDEC separates error flagging, sentence localization and correction. It replaces the old guessed leaderboard with exact attributed results, construction methods, comparison subsets and conditional metric denominators.

Exit test: explain how a model can score well on generated corrections while missing many errors, and why a JSON schema or human reviewer alone does not establish safety. The exercise is benchmark reading, not treatment advice, a deployed clinical tool or authorization to upload patient records.

Part 7 · Statistical grounding, optional but substantial

Begin with the ML mental model, then linear regression, logistic regression, regularization, bias and variance, and model selection/CV. These explain the experiment underneath a training or evaluation claim.

For the university course itself, use Stanford CS229's official course materials. Prepare with linear algebra, probability, calculus, optimization and generalization; a short overview cannot replace practice with those foundations.

Exit test: define train/validation/test roles, choose a temporal or grouped split when deployment requires it, fit preprocessing inside each training fold and compare against a baseline. Return here whenever “accuracy improved” lacks a denominator or valid holdout.

Part 8 · Shipping and operating boundaries

Foundry platform concepts and reference workflow distinguishes resource/project/deployment identities and SDK generations, grounding, authorization, evaluation, observability and cost. A platform page is a source-supported design reference, not proof that a model deployment ran.

Exit test: record the model/deployment/API version, permitted data route, identity roles, evaluation set, timeout/retry policy, rollback path and budget. Grounding a response is not permission to reveal its source. Do not create resources merely to check off a reading route.

How to read this series

For an orientation session, read parts1 and3, then sketch the requirements in part8. That yields an architecture discussion—not “deployed in an hour.” For the mechanics route, read1→2→3 and calculate the examples. For application work, add4, then choose audio5 or clinical6, and use7 to critique your evaluation. Reading times are estimates; use exit tests rather than a fixed deadline.

Use the links here as the explicit route. Automatic previous/next navigation can follow each destination's own plan and may not reproduce this cross-topic order. Keep a small notebook with the example you solved, one negative case and one unanswered question; opening a page alone is not completion.

What this series does not replace

• Training and fine-tuning from scratch: the full 80-session DL map covers math, models, pretraining, adaptation, serving and capstones. A small API app is not a substitute for that breadth.
• Agents and tool execution: agents/tools develops the controller and permission boundary. It is an adjacent path, not a reason to omit tools from an application's threat model.
• Evaluation in depth: LLM evaluation and the clinical/RAG case studies distinguish judges, gold sets, regression tests and real outcomes. Passing a toy fixture does not certify a product.

The conceptual sources include Attention Is All You Need, RAG, and Microsoft's security-filter pattern. Use this map to choose a learning route; each linked article explains its sources, examples and execution limits. No API, deployment, clinical inference or paid service was executed for this map.