Page 54
12 articlesLLM-Powered Data Pipelines: The ETL Tier Nobody Benchmarks
Most LLM benchmarks measure chatbot quality. But the bulk of enterprise LLM spend is going into batch pipelines — and almost nobody is measuring whether those pipelines actually work.
LLM Vendor Lock-In Is a Spectrum, Not a Binary
Not all LLM dependencies are created equal. Some are acceptable engineering tradeoffs; others are technical debt from day one. Here's how to tell them apart across six distinct lock-in layers.
Long-Session Context Degradation: How Multi-Turn Conversations Go Stale
Sessions beyond 50 turns accumulate contradictions, user intent drift, and sycophancy loops. Here's the engineering playbook for detecting degradation and keeping long conversations useful.
The Long-Tail Coverage Problem: Why Your AI System Fails Where It Matters Most
Aggregate metrics like accuracy and F1 can look great while your AI system silently fails on the minority inputs that matter most. How to detect, measure, and fix long-tail coverage gaps before users find them.
LoRA Adapter Composition in Production: Running Multiple Fine-Tuned Skills Without Model Wars
Teams build separate LoRA adapters for tone, format, domain knowledge, and safety — then hit conflicts when composing them. Here's how to detect interference, choose the right merge strategy, and serve mixed adapters per-request without reloading weights.
The Mental Model Shift That Separates Good AI Engineers from the Rest
The transition from deterministic to stochastic systems trips up strong engineers. Here are the mental models, debugging intuitions, and practices that actually separate experienced AI engineers from everyone else.
Model Deprecation Is a Production Incident Waiting to Happen
LLM providers deprecate models on 6–12 month windows, but most teams treat migration as a backlog item—until it becomes a 3 AM outage. Here's the operational playbook to make model upgrades boring.
Multi-Tenant AI Systems: Isolation, Customization, and Cost Attribution at Scale
How to serve multiple customers from shared AI infrastructure without leaking data, creating noisy neighbors, or losing track of who's spending what.
Multi-Modal Agents in Production: What Text-Only Evals Never Catch
Adding vision and document inputs to agent pipelines introduces failure modes that text-only evals never surface. Here's what practitioners are running into and how to build evals that catch it.
Multimodal AI in Production: The Gap Between Benchmarks and Reality
Vision and audio models look impressive in demos. In production, they face latency penalties, grounding failures, and extraction inconsistencies that most benchmark scores hide entirely.
The 90% Reliability Wall: Why AI Features Plateau and What to Do About It
Why AI features stall around 90% reliability, how to diagnose reducible vs. irreducible error, and the product-architecture decisions that let you ship honest value.
On-Call for Stochastic Systems: Why Your AI Runbook Needs a Rewrite
Traditional incident response assumes reproducible failures. LLM-powered systems don't. Here's how to rewrite your alerting schema, triage decision tree, and post-mortem template for non-deterministic AI.
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.
