Learn
AI engineering, taught from real code
A short course for CTOs, developers and technical readers. Every concept is explained on its own terms first, then shown working — or honestly not working — in one of our open-source products, and placed next to the tools the rest of the ecosystem uses.
Every lesson separates three kinds of claim:
- In the product — exists in the repository; the file is cited.
- Concept — holds with or without our products.
- Ecosystem — external tools and practices; they change fast, verify before deciding.
Terms with a dotted underline show a short definition on hover, focus or tap.
Case studies
- AdaptaSelf-hosted RAG and QLoRA fine-tuning behind an evaluation gate. Unaudited prototype.Source · Apache-2.0 (opens in new window)
- docketGoverned runtime for coding-agent teams, one chokepoint for every tool call. Beta.Source · Apache-2.0 (opens in new window)
- TackSelf-hosted project board that dispatches work to coding agents through durable runners. Public beta.Source · MIT (opens in new window)
Track 1 · 8 lessons
ML and LLM foundations
What a model is, how it learns, and how to give it knowledge and behavior without losing control. Case study: Adapta.
- 01A short history of AIFrom the perceptron to agents: the winters, the bets that paid off, and why each wave changed what engineers build.12 min
- 02Machine learning in one lessonData, training, inference, generalization and overfitting: the vocabulary everything else builds on.10 minAdapta
- 03LLMs as software componentsTokens, context, temperature and structured output, and why a 200 OK no longer means the answer is right.12 minAdapta · docket
- 04Models, open weights and Hugging FaceClosed and open-weight model families, licenses and model cards, and the Hugging Face Hub as the place models are shared.13 minAdapta
- 05Embeddings and RAGChunking, vector and keyword search, rank fusion, reranking and citations: giving a model knowledge it was never trained on.14 minAdapta
- 06Fine-tuning, LoRA and QLoRAWhen to fine-tune instead of prompting or retrieving, and how adapters and quantization make it fit on one GPU.14 minAdapta
- 07Evaluating modelsHeld-out data, baselines, metrics and eval gates: how to know a model got better instead of hoping it did.12 minAdapta
- 08Serving modelsLocal versus hosted inference, OpenAI-compatible endpoints, quantized weights, tenancy and the real cost drivers.11 minAdapta · docket
Track 2 · 10 lessons
Agents and AI engineering
From calling a model to governing one that acts: loops, tools, policy, isolation, observability and security. Case study: docket.
- 09From LLM app to agentLLM apps, workflows and agents differ in who picks the next step. The agent loop, and why it is a security surface.12 mindocket
- 10Tool calling and harnessesThe model asks, your program executes. Tool schemas, argument validation, and the runtime that owns that boundary.13 mindocket
- 11Coding agents: harnesses and skillsClaude Code, Codex, OpenCode and friends; instruction files, skills, subagents and hooks, and how to drive several harnesses from one system.13 mindocket · Tack
- 12Governance and human-in-the-loopPolicy as code, deny-ask-allow, risk classification and approvals that a person can actually give.13 mindocket
- 13Sandboxing and agent securityPrompt injection, the lethal trifecta, least privilege and isolation levels, from worktrees to microVMs.14 mindocket · Tack
- 14Context, routing and costContext engineering, compaction, model routing by role, budgets, and why measured tokens beat estimated dollars.11 mindocket
- 15Model gatewaysOne endpoint in front of many providers: keys, fallbacks, rate limits, caching, cost tracking and where a gateway stops helping.11 mindocket
- 16Observability, audit and reliabilityTraces, hash-chained audit logs, retries, timeouts and cancellation: evidence that survives after the turn ends.13 mindocket · Tack
- 17Testing and evals for agentsFake models, architectural tests, golden files and eval suites: testing a system whose core cannot be read as code.12 mindocket
- 18MCP, the Model Context ProtocolWhat MCP standardizes, what it does not, and how to expose or consume tools without giving up control.10 mindocket · Tack
Track 3 · 4 lessons
Orchestration in production
Teams of agents, durable execution, and the decisions a CTO has to make before adopting any of it. Case studies: Tack and docket.
- 19Multi-agent systemsRoles, pipelines and independent verification, and when one agent is the better answer.12 mindocket · Tack
- 20Durable agent executionPull-based runners, leases, fencing tokens and ambiguous outcomes: running agents that can crash without running twice.13 minTack
- 21From vibe coding to the lights-out factoryThe spectrum from autocomplete to unattended software factories, and the specs, tests and gates each step demands.13 minTack · docket
- 22Adopting AI engineering: a CTO guideArchitectures by maturity level, build versus buy, a decision table, anti-patterns and a skills roadmap.14 mindocket · Tack · Adapta
Track 4 · 2 lessons
In practice
Documented, public uses of AI engineering, and one system built end to end with the three case studies.
- 23Applied use casesFive worked scenarios — support, coding, internal search, documents and operations — with what you build, what breaks first and what to measure.14 min
- 24Implementation walkthroughA support assistant built end to end: knowledge with Adapta, a governed agent with docket, work tracked in Tack, and a deployed chat widget.16 minAdapta · docket · Tack
Reference