Model I/O
Raw API chatbot with streaming, typed output, a tool loop, retries, and cost logs.
Applied path · senior engineer edition
From shipping software to shipping AI systems.
A complete, text-first twelve-week field course: concepts, architecture diagrams, runnable labs, evaluation gates, debugging notes, and interview evidence—without a video syllabus.
- Cadence
- 8–10 hrs / week
- Format
- Read · draw · build
- Outcome
- 4 portfolio proofs
- Baseline
- Senior SWE
The operating model
You do not need to relearn software engineering. Add the probabilistic layer, then make it observable, evaluable, and safe.
Six phases, four proofs
Every phase ends with something inspectable. The strongest portfolio is not four chat UIs; it is four measured systems with failure analysis.
Retrieval
Document QA with dense + lexical retrieval, reranking, citations, and a held-out test set.
Orchestration
Bounded stateful agent with explicit transitions, approval gates, tracing, and MCP.
Adaptation
Small LoRA/QLoRA experiment with a data card and before/after behavioral eval.
Reliability
Golden dataset, regression gate, adversarial suite, and failure taxonomy.
Production
Load-tested inference path, routing policy, cost model, runbook, and design narrative.
Rules of the course
Measure before frameworking.
Begin with raw model calls, a tiny dataset, and a visible trace. Frameworks become valuable only when the repetition is understood.
Separate retrieval from generation.
A fluent wrong answer may come from bad search, bad context use, or a bad source. Score the stages independently.
Bound agent freedom.
Use deterministic code for rules, the model for judgment, and explicit approval for consequential actions.
Evidence beats feature count.
A graph of quality, latency, and cost plus three analyzed failures is stronger than another polished chatbot shell.
Retrieval practice
Knowledge check
Try to answer aloud, then reveal. The point is to rehearse engineering decisions, not terminology.
Evidence base
Sources & further reading
The lessons are self-contained; these primary papers and official docs support the claims and let you inspect the originals when you want to go deeper.