Practical insights
on software and AI

IKRC team members hold Anthropic's architecture credential. An invoice approval example shows where that knowledge matters: application permissions and recovery when an action's result is uncertain.

Microsoft states it in an Important callout: MCP server authorization defines access to the server and does not reach individual tools. The protocol makes authorization optional, any connected caller can list every tool and read every description, and the token you issue must not be forwarded to the systems behind. What holds the line is the tool calling the same service method your screens call, so that one rule change covers both and neither drifts.

Azure AI Search offers four ways to do document-level access control, and only security filters are generally available today. Whichever one you pick, the index stores its own copy of the permissions, captured the last time an indexer looked. Changes at the source do not reach it on their own: a parent-scope change in SharePoint and any permission change in ADLS Gen2 both need an explicit resync, and until one runs the index keeps answering from what it captured.

CREATE EXTERNAL MODEL and AI_GENERATE_EMBEDDINGS turn the embedding step into a statement over a table. What decides whether it survives contact with production sits around the edges: a database scoped credential whose name has to be the endpoint URL, a retry_count between 0 and 10 that a query can override, an extended event that is the only place the REST status code shows up, and a chunking function whose one chunk type counts characters.

AddChatClient goes into the same startup file as the repository and the mail sender, resolved by the same dependency injection everything else uses. The part worth deciding deliberately is the order: a cache hit returns without forwarding the request along the pipeline, so nothing registered after the cache runs, and the lookup matches on the whole submitted chat history rather than on the last thing a user typed.

The scope of an agent is its tool list, not the sentence in its prompt. Its failure path is a retry loop with published defaults, and those defaults decide whether the support desk sees the account number or a generic error three times over. Done is a row that changed, not a satisfied user.

A build-versus-buy decision starts with the workflow a customer would notice, then runs two diagnostics against the incumbent product: how its configured logic is developed and released, and whether one complete slice of records, associations, history, and files can be reconstructed somewhere else.

Most disappointing AI assistants are not model problems. They are retrieval problems: the wrong copy of the record, no delete path into the index, a service account that reads more than the user can, and no log of what was actually retrieved. Fix those four before adding the model.

SQL Server 2025 stores embeddings natively and searches them with VECTOR_DISTANCE, both generally available. The approximate DiskANN index is preview and carries a deployment constraint that catches teams with a DacPac pipeline. For most business workloads the exact-search path is the answer.

Support for .NET 8 and .NET 9 ends on November 10, 2026. Before committing to an upgrade date, check package compatibility, vendor commitments, deployment type, runtime, hosting, and build agents.

How to design multi-tenant business software with tenant isolation, permissions, reporting, billing, and support visibility built in from the start.

Why companies should consider MCP servers that let customers book appointments, reorder products, check status, and ask for help from their AI assistant.

How AI-integrated applications help law firms collect diligence material, evaluate responses, flag gaps, and suggest follow-up questions for attorney review.

How custom AI knowledge bases preserve institutional knowledge, keep answers current, and protect information by role, account, department, and audience.

What WebMCP means for websites, how browser agents discover page tools, and how companies can prepare one high-value flow at a time.

A practical look at when custom software platforms should expose MCP servers, what to expose first, and how to keep AI access controlled.

How to build API integrations that keep working after launch, with support screens, ownership rules, retries, reconciliation, and real workflow handling.

How to evaluate a C# and .NET development company when the software carries real workflows, data, integrations, and long-term support needs.

A practical guide to modernizing legacy applications by reducing operational risk, preserving business rules, and replacing fragile parts in the right order.