AI Engineer Sandbox Event

We replaced a frontier-model dependency with a fine-tuned 8B open model and cut inference cost 14x. A practical walkthrough of dataset curation, LoRA vs full fine-tuning, evaluation, and the serving setup — every number included.

A hands-on workshop built from real agent incident reviews. Bring a laptop: we will reproduce four production failure modes — runaway loops, silent tool errors, stale context, and compounding hallucinations — and fix each one with guardrails you can take home.

Skyforge Cloud's DevRel lead shows what teams are shipping on Skyforge's inference platform, with a live build of an agentic workflow from zero to deployed in twenty minutes.

Everyone says they run evals; almost nobody catches regressions before users do. This talk walks through the evaluation harness we built after three silent quality incidents: golden sets that stay fresh, LLM-judges you can actually trust, and the CI wiring that blocks a bad prompt from shipping.

The RAG stack of 2024 is a liability in 2026. This talk covers what replaced it in our systems: hybrid retrieval with learned rerankers, query planning, and retrieval evals that correlate with user outcomes — with benchmarks from two production migrations.