Engineering field notes
How things are set up. What made them work.
And what was learned along the way.
A place for the details behind the work: practical walkthroughs, small experiments, and explanations to come back to.
-
006 / DATAKEEPERHDN / BEDROCK
Fine-tuning Llama 3.2 1B for grounded HDN drafts on Amazon Bedrock
Fine-tuned Llama 3.2 1B Instruct locally on customer-approved HDN packets, imported the weights to Bedrock, and rejected any draft whose names or amounts were not in the packet.
READ THE NOTE -
005 / SUPPLY CHAINAZURE DEVOPS
CycloneDX SBOM and container attestation in Azure Pipelines
Generate the bill from the image Azure Pipelines built, mark agent-added components, and attest it to the digest.
READ THE NOTE -
004 / AGENTSTYPESAFE
Typed context routing for AI-assisted development with TypeSafe Jev
One typed call before the agent opens the repo. Intent, bundles, evidence. A label does not authorize the work.
READ THE NOTE -
003 / SPEECHVITS
Training a multi-speaker VITS speech model.
A VITS voice in PyTorch, trained so a client’s answer robot could speak. The cut, the pitch list, and the sample rate that had to match.
READ THE NOTE -
002 / AIQUESTION BOT
Production RAG chatbot: retrieval, context limits, and evaluation.
How I keep the Ramazan Demir chatbot inside his own writing: the chunk budget, the query guard, and the refusal.
READ THE NOTE -
001 / PLATFORMSOPENSEARCH
Centralized AWS logging with OpenSearch, Cognito, and Terraform
Cross-account logs into one OpenSearch domain, Cognito in front of the dashboards, written up from the Datakeeper logging setup.
READ THE NOTE