AI
insights
Practical notes on AI, newest first.
Reliable Structured Output from Enterprise AI: Schema Constraints, Validation, and Repair
A practical architecture for producing AI-generated data that can safely enter ERP, CRM, databases, and automated workflows.
Detecting and Masking Sensitive Data Before Model Calls
A practical pipeline for classifying, detecting, masking, and auditing sensitive data before it reaches an AI model.
Recovering Interrupted Streaming Responses in Enterprise AI
A practical architecture for cancellation, resumable delivery, idempotent retries, and reliable state recovery in AI streaming systems.
What Should AI Remember? Retention and Deletion Design for Enterprise Conversations
Enterprise AI needs purpose-based memory, explicit retention periods, and deletion workflows that remove source data and every derived copy.
Combining Keyword, Semantic, and Reranking for Enterprise Search
A practical architecture for combining hybrid retrieval, reranking, access control, and measurable relevance in enterprise search.
Confirmation and Undo Patterns for AI-Executed Tasks
A practical design guide for confirmations, compensating actions, and execution receipts in enterprise AI workflows.
Consistent Terms and Citations in Bilingual AI Assistants
A practical architecture for keeping terminology, evidence, and citations aligned across Chinese and English AI responses.
When Enterprise Terms Go Wrong: Fine-Tuning or Better Retrieval?
A practical framework for choosing retrieval, fine-tuning, or a hybrid approach when AI repeatedly misinterprets enterprise terminology.
Routing Enterprise AI Queries by Data Freshness
A practical architecture for using timestamps, intent classification, and source routing to keep enterprise AI answers current.
Balancing Offline and Online Evaluation for Enterprise AI
Offline evaluation establishes quality and safety gates, while online evaluation verifies real-world value and feeds new evidence into future releases.
Preserving Numbers, Terms, and Ownership in AI Summaries
A practical engineering approach to keeping critical figures, contractual conditions, and accountable owners intact in AI-generated summaries.
Choosing Enterprise Workloads for Small Language Models
A practical framework for selecting small-model workloads based on scope, latency, cost, deployment constraints, and failure risk.
Have a project in mind?
Tell us your industry, current systems and budget range. Free 30-minute consultation.