Page 119
12 articlesYour Database Schema Is Your Agent's Mental Model
Poorly normalized schemas cause AI agents to hallucinate joins, misread relationships, and chain unnecessary tool calls. Here's how to design a schema layer that your agent can actually reason about.
Embedding Models in Production: Selection, Versioning, and the Index Drift Problem
Picking the wrong embedding model—or failing to manage upgrades—silently kills RAG retrieval quality. A practical guide to model selection beyond MTEB scores, detecting index drift, and zero-downtime versioning strategies.
Feature Flags for AI: Progressive Delivery of LLM-Powered Features
Rolling out LLM-powered features requires more than traditional feature flags. A guide to prompt variant management, the three-tier metric stack, cohort consistency for multi-turn sessions, silent degradation detection, and rollback strategies that actually work.
Fine-Tuning Economics: The Real Cost Calculation Before You Commit
Most teams underestimate fine-tuning costs by 3–5x because they only budget the training run. Here's the complete cost model — data curation, failed experiments, deployment, maintenance — and a decision framework for when LoRA/PEFT actually beats months of prompt engineering.
GraphRAG in Production: When Vector Search Hits Its Ceiling
Vector search fails predictably on multi-hop reasoning queries. GraphRAG addresses that gap — but introduces a different cost structure, failure modes, and maintenance burden that most teams underestimate.
Knowledge Distillation Economics: When Compressing a Frontier Model Actually Pays Off
The real cost math behind compressing frontier models into specialized smaller ones — when distillation beats fine-tuning, when it doesn't, and the failure mode where students inherit their teacher's confident wrongness.
Releasing AI Features Without Breaking Production: Shadow Mode, Canary Deployments, and A/B Testing for LLMs
Shipping a new model version or prompt change to production carries risks that standard deployment processes don't catch. Here's how shadow mode, canary deployments, and A/B testing work together for safe LLM releases.
Where Production LLM Pipelines Leak User Data: PII, Residency, and the Compliance Patterns That Hold Up
Most LLM data leaks don't come from the model — they come from unredacted RAG chunks, verbatim prompt logs, and injectable retrieval pipelines. A practical guide to PII handling, data residency routing, and compliance logging for production AI systems.
Long-Context Models vs. RAG: When the 1M-Token Window Is the Wrong Tool
Why dumping your entire knowledge base into a 1M-token context window fails in production — the latency, cost, and accuracy tradeoffs that make RAG the right default for most retrieval workloads, and a five-factor decision framework for when long context actually wins.
Why Your Agent Harness Should Be Stateless: Decoupling Brain from Hands in Production
Tight coupling between AI agent harnesses and sandboxes kills reliability, scalability, and security. Here's the architectural pattern that fixes it: external session logs, stateless harnesses, and isolated sandboxes.
The Model Upgrade Trap: How Foundation Model Updates Silently Break Production Systems
Foundation model updates silently break production systems through behavioral drift, changed refusal patterns, and JSON serialization inconsistencies — a practical guide to detection and safe migration.
Multi-Tenant LLM API Infrastructure: What Breaks at Scale
A production API gateway in front of LLM providers solves cost attribution and rate limit contention — but the hierarchical isolation model, token-aware limits, failover patterns, and KV cache security create complexity most teams underestimate until they're already burned.
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.
