Page 15
12 articlesThe Avatar in the Conference Call: Engineering Real-Time Talking-Head AI for Video Meetings
A 45ms audio-video offset is the threshold where humans flag a talking-head AI as fake. Inside the real-time engineering — viseme schedules, audio master clocks, and the failure modes that never show up in offline eval.
The Reply-All That Wasn't: Agent Outbound Fan-Out Hazards
Agents have no muscle memory for Reply All. When a send tool's recipient field accepts a distribution list indistinguishably from an individual, the planner picks the loudest door. Four practices to keep blast radius bounded.
Retiring an Agent Tool the Planner Learned to Depend On
Removing a function from your agent's tool catalog isn't a syntactic change — it's a behavioral migration. The staged retirement pattern that prevents fallback hallucinations and silent regressions.
Runtime Prompt Hot-Reload: Why Your Prompts Shouldn't Be Locked Behind a Build
Coupling prompt changes to your deploy pipeline is a self-imposed constraint. The runtime hot-reload pattern, its safety primitives, and the failure modes nobody plans for.
The Shadow AI Governance Problem: Why Banning Personal AI Accounts Makes Security Worse
Blocking employees from personal ChatGPT or Claude accounts doesn't stop AI use — it makes it invisible. Here's how to survey shadow AI, build sanctioned channels, and avoid governance theater.
The SIEM Bill Your AI Feature Forgot to Include
Shipping an AI feature multiplies your audit-log volume 10–50x. The SIEM renewal arrives later, with broken detection rules and a legal-hold problem nobody scoped.
Silent Quantization: Why the Model You Pay For Today Isn't the Model You Paid For Last Quarter
Vendors quietly re-host the same model name on cheaper precision tiers as inference economics tighten. Here is why version strings stop being a contract — and the probe set, routing layer, and SLA clause that replace them.
Skills as Modules: When Your Agent Stack Needs an Import System
Agent runtimes are reinventing import systems badly — name resolution, version pinning, dependency graphs, and conflict detection are unsolved problems hiding under the skills ecosystem.
When to Skip Real-Time LLM Inference: The Production Case for Async Batch Pipelines
Most LLM workloads are batch-compatible, but teams default to synchronous calls because that's what the API makes easy. Here's the break-even analysis and the feature categories where async beats streaming on cost and UX.
Snapshot Tests Lie When Your Model Is Stochastic
Snapshot tests assume same input plus same code equals same output. Once an LLM call is in the loop, that contract breaks and the suite quietly turns into a rubber stamp. Here is the testing taxonomy that replaces it.
The Tail-Tolerant Retry Policy Your LLM Gateway Doesn't Have
Microservice retry defaults applied to an 8-second LLM call inflate P99, burn tokens during provider incidents, and hide a customer-visible latency cliff that the gateway dashboard never shows.
Why Token Forecasts Drift After Launch — and How to Catch the Spike Before Finance Does
Pre-launch cost models assume a synthetic traffic mix. Post-launch reality shifts the moment the feature actually works. The bill is the worst possible detector — here's how to catch the drift in real time.
About Tian Pan
I'm Tian Pan, an engineer-founder focused on agentic engineering — building autonomous AI systems and scaling engineering teams. I write practical guides on system design, technical leadership, and shipping with AI agents. Previously an early engineer at Uber, Brex, and IoTeX.
