Building and Operating Conviro: The Work Around the Model
Engineering lessons from building Conviro: tenant boundaries, retrieval quality, provider routing, tool permissions, and operating responsibility.
AI Engineering 4 min read
Blog
Notes from building and operating AI systems, commerce platforms and the measurement around them.
AI Coding Agents Need a Verifiable Deployment Boundary
Connect the code reviewed, the code tested, and the artifact deployed so an AI-assisted workflow cannot confuse completed work with a safe release.
Agentic Systems 3 min read
LLM Evals Should Exercise the Path You Actually Ship
Evaluate routing, retrieval, fallback, validation, and handoff together, then use diagnostic checks to explain where failures originate.
AI Engineering 3 min read
A Tool-Using Agent Needs Server-Side Authorization
Treat model-generated tool calls as proposed actions. Validate permissions, scope, arguments, budgets, and approval at execution time.
Agentic Systems 3 min read
Streaming LLM Responses: Cancellation and Backpressure Matter
A reliable streaming response needs explicit lifecycle events, bounded buffering, upstream cancellation, and a clear policy for partial answers.
Backend Architecture 3 min read
The Transactional Outbox Closes One Gap, Not Every Gap
Use an outbox to connect database changes with event publication, then address duplicate delivery, ordering, and external side effects separately.
Backend Architecture 3 min read
LLM Budgets Must Survive Concurrent Requests
Turn a usage dashboard into an enforceable spending policy with atomic reservations, bounded calls, reconciliation, and tenant-aware accounting.
AI Engineering 3 min read
Multi-Provider LLM Routing Is a Policy Problem
A useful model router combines capabilities, permissions, region requirements, budgets, and failure behavior behind an explicit contract.
AI Engineering 3 min read
Multilingual Support Needs More Than Translating the Answer
Keep language, market, currency, and policy scope separate so a multilingual assistant retrieves and communicates the right information.
AI Engineering 3 min read
Checking an AI Answer Before a Customer Sees It
Separate fluent generation from evidence checks for prices, product identifiers, links, and policy claims before releasing an AI response.
AI Engineering 3 min read
Hybrid Search with PostgreSQL: When Keywords Still Matter
Combine semantic retrieval with lexical search while preserving identifiers, access filters, and a measurable baseline for ranking quality.
AI Engineering 3 min read