Tag archive
How I picked an LLM for food photo recognition in soba: a benchmark, a live A/B that cut scan cost to $0.005, and the prompt work that fixes portion weight.
If every decision is hard-coded, the LLM is just an expensive text formatter. Cogweald draws a clear...
TypeSafe’s early-access Jev turns predefined choices, scores and probabilities into fast API calls, but its proprietary architecture and reliability claims remain unverified outside company-run evaluations.
Schemas, validation, and retries: making model output machine-readable every single time.
Hugging Face lifted a 350M model's structured-output score with 100 GRPO steps and no labels. The trick isn't the model, it's three reward functions y
Day-one 0731 measurements: $0.14/$0.28 list, a 1,024-token cache page, and a defect: thinking plus strict json_schema corrupted integers in 8 of 13 runs.
Originally published at norvik.tech Introduction Explore the technical aspects of...
A model that occasionally breaks a JSON schema under prompting alone can be fine-tuned to be far more reliable. When this is worth doing over prompt e
Short answer: use chat completions with a strict JSON schema for small-scale support-ticket...
Short answer: Put moderation behind one OpenAI-compatible chat-completions contract, require a...
Short answer: build moderation around one validated JSON decision and an OpenAI-compatible...
Two production failure modes that break LLM JSON parsing—and how to detect and degrade gracefully. Includes exact code for 400 retries and a fence-safe parser.