Topic
AI Engineering
Agents, RAG and LLM features that survive production: contracts, evaluation, authority, cost and recovery.
Human Approval After a Restart: One Reviewed Proposal, One Local Effect
Bind approval to the exact draft, evidence and policy, resume a pending review from an Agent Framework checkpoint in a new process, and recover a lost receipt without a second effect.
When an Agent Coordinator Has a Real Choice
Drop selector calls where the dependency graph already decides, keep a checked selector only at a real fork, and charge every attempt so a fluent loop that makes no progress still stops.
RAG That Keeps Its Evidence: Eligibility, Coverage and Licensed Claims
Filter before Top K, check that the selected set covers every required fact, accept only sentences their citations license, and carry an evidence identity into the next stage.
Local Models and Polish Routing: Evaluate the Errors, Not the Average
A local model changes where inference runs, not what its scores mean. Evaluate English and Polish pairs, separate agreement from correctness, and see why a margin threshold misleads.
Azure Decision APIs: Root, Deployment, Operation and Token Audience
Microsoft-Decision-1 and Azure OpenAI Decisions use different paths, serializers and token audiences. Bind each to one validated C# profile and keep the token inside the deadline.
Native Decision API Clients in C#: Wire Shapes and Stalled Bodies
Build one C# client for System One and OpenAI Decisions that maps answers by meaning, parses strictly, owns one deadline past the headers and reports every failure as a distinct outcome.
Decision APIs: Derive the Threshold From the Cost of Each Error
A probability is not an action. Derive automation and review thresholds from error costs, model imperfect reviewers, rank a scarce queue and check what calibration must hold.
Designing a Decision Contract That Preserves Meaning in C#
Write the rubric first, give choice, predicate and score their own C# types, and reject what a JSON schema cannot see: missing candidates, inconsistent means and duplicate keys.
What a Decision Probability Actually Computes: Logits to Labels
The same {"department":"technical"} can come from generation, a classifier or label scoring. Work through softmax, candidate sets, multi-token labels and schema masks to see what 0.88 means.
Decision APIs in .NET: One Judgment Inside a Working Support Desk
Give a model one bounded question, which team investigates first, and let C# policy decide the action. A small Agent Framework workflow shows six tickets stopping for six reasons.
From AI Draft to Accepted Work Order: One Service, Real Review
A complete-looking AI draft can pass every validator and still overstate its evidence. Route forms and IChatClient through one service and bind approval to revision and evidence.
Move ONNX Inference Behind HTTP Without Changing the Computation
A matching signature and a unit-length vector prove little. Pin the MiniLM pipeline, fix three .NET tokenizer differences, pool with the mask and know what a tolerance guarantees.