Dev3Pack AI Engineering Bootcamp

14.09 → 02.10.2026 · Certificate earned

Certificate of completion · October 2026Verify ↗

17/17 chapters100 %

Summary

Three weeks to build a source-grounded research assistant: it answers developer questions from six documents, cites the document it used, and refuses when nothing backs the answer.

Bootcamp completed: 5,100/5,100 points on the track, final assignment at 100%, certificate earned.

17 published · 0 coming soon

Session 01Published
Setting up the assistantA tags filter, two rejected changes, and one rule: only the diff counts.Read the chapter →
BAA 101 concept
- Verify the effect, not the claim
Session 02Published
The model adapterOne door to every model, an outage that becomes a refusal, and a timeout set by measurement.Read the chapter →
BAA 101 concept
- Refusal is a state
Session 03Published
Structured outputsA strict parser, a single retry, and a golden set tested before it's labeled.Read the chapter →
BAA 101 concepts
Session 04Published
Bounded toolsA tool validates before acting, and what it returns is data, never an order.Read the chapter →
BAA 101 concepts
- The tool as the model sees it
- Data is never an order
Session 05Published
A deterministic mini-agentA loop that always stops, and says why: four exits, a budget counted before the call.Read the chapter →
BAA 101 concept
- The agent loop and its exits
Session 06Published
Retrieval baselineA word-based search, a loader that refuses, and verdicts read in the results, not in the score.Read the chapter →
BAA 101 concepts
- Refusal is a state
- The right page must arrive
Session 07Published
Retrieval and grounding metricsA fix at 100% that breaks unseen cases, and a cheating fake model that scores 50%.Read the chapter →
BAA 101 concepts
- The right page must arrive
- Evaluate without lying to yourself
Session 08Published
Loops and graphsChain, loop or reflection: count the calls, and a table where a refusal never becomes an answer.Read the chapter →
BAA 101 concept
- The agent loop and its exits
Session 09Published
Tracing and evaluationThe trace says where the failure is, and secrets go in masked, never deleted.Read the chapter →
BAA 101 concept
- Read the tool call, not the answer
Session 10Published
Skills and ADRsA buyer that refuses before paying, an ADR with its reversal threshold, and a skill judged on evidence.Read the chapter →
BAA 101 concept
- Data is never an order
Session 11Published
State and memoryMemory filed by owner, capped, with a policy written before the first record.Read the chapter →
BAA 101 concept
- Memory has an owner
Session 12Published
MCP architectureSixteen advertised tools, two that move money: reading an MCP server like a prospectus.Read the chapter →
BAA 101 concept
- The tool as the model sees it
Session 13Published
Securing an MCP serverDecide before opening the connection, and tick an attestation only if it's true.Read the chapter →
BAA 101 concept
- The tool as the model sees it
Session 14Published
Deploying and operatingA smoke test must be able to fail, and the rollback sentence is written before the outage.Read the chapter →
cap01Published
The final's notebookA green evaluation on a fake model proves the plumbing, not reliability.Read the chapter →
BAA 101 concept
- Evaluate without lying to yourself
FinalPublished
Final assignmentFour prompt tweaks with no stable effect, then one idea: the model chooses, the code copies. 15/15.Read the chapter →
BAA 101 concepts
GeckoPublished
Gecko capstoneA buyer that pins the request, checks every field, and signs only if everything matches: a real purchase on devnet.Read the chapter →
BAA 101 concepts

GitHub repo See BAA 101

Summary

Contents