Posted inAI AWS Machine Learning
Day 10: Project: Internal Docs Chat With Citations
Ship an internal docs chat on Bedrock + OpenSearch or pgvector that refuses uncited answers — a project lab you can demo.
Day 35: Idempotent Tool Calls Against DynamoDB
Use DynamoDB conditional writes and client idempotency tokens so agent retries never double-charge, double-ticket, or double-deploy.
Posted inAI Machine Learning
Day 9: Eval Harness Day One
Golden questions, faithfulness, hit-rate, and a CI job that fails the build — evals on day one, not after the outage.
Day 34: EventBridge as the Agent’s Async Backbone
Offload minute-scale agent tools through EventBridge so sessions stay interactive — correlation IDs, status tools, and no blocked Bedrock turns.
Posted inAI
Day 8: Structured Outputs That Compile
JSON Schema, tool-choice, and validators so model output compiles — inventing fields is a bug, not creativity.
Posted inAI AWS System Design
Day 33: Agents on Step Functions, Not Infinite Loops
Run agents as Step Functions graphs with human approval states, timeouts, and compensations — not while(true) ReAct loops in Lambda.
Posted inAI System Design
Day 7: Prompt Contracts: System, Developer, User, Tool
Separate system, developer, user, and tool prompt layers like IAM roles — privileges leak when everything is one blob.
Posted inAI
Day 6: Chunking Strategies for Code, Tickets, and Runbooks
Fixed, recursive, AST, and late chunking each win on different corpora — pick by code, tickets, and runbooks, not blog defaults.
Day 32: Bedrock Converse API Tool Choice in Production
Pin Converse toolChoice (auto vs any vs specific tool), version prompts in Prompt Management, and stop agents from freelancing tool calls in production.
Posted inAI AWS Machine Learning
Day 5: Your First RAG Pipeline That You Can Debug
Build a debuggable RAG loop — ingest, chunk, retrieve, generate — with traces, citations, and an explicit failure budget.