TESTSStructured outputs vs. plain prompts: when constrained decoding is worth it
Schema-constrained decoding guarantees valid JSON. It does not guarantee correct data, and it can cost quality. How to choose and how to test.
PROMPTSSystem prompt templates for production: six copy-ready examples
Support, extraction, grounded Q&A, code review, summaries and classification. Templates you can adapt, with the failure each line prevents.
AGENTSStep budgets and stopping rules for AI agents
Every agent needs to know when to stop. Budgets, loop detection, progress checks and graceful exits that keep agents useful and affordable.
PROMPTSPrompting for citations that actually check out
Source IDs, quote-then-answer, automatic quote checks and a real "not found" path. How to make every claim traceable and verify it in code.
MODELSSmall models, big jobs: where tiny models quietly win
Classification, extraction, routing and on-device work. The jobs where a small model matches a big one at a fraction of the cost, and how to prove it.
PAPERSMixture of experts, without the hype
Many parameters, few used per token. How MoE routing works, what "active parameters" really means, and the trade-offs for anyone running or buying these models.
BENCHMARKSDesigning an LLM latency benchmark that matches what users feel
Time to first token, tokens per second and p95 under real load. A method for measuring LLM speed that predicts user experience, not lab conditions.
AGENTSHuman-in-the-loop patterns that don’t annoy the human
Approve everything and people stop reading. Risk-based approvals, batching, good review screens and escalation patterns that keep humans in control.
PROMPTSDecomposition prompts for long, messy tasks
One giant prompt fails in ways you cannot see. Chaining, map-reduce and plan-then-execute turn long tasks into steps you can test and fix.