In Ramzor, Hefer survives a high-tech job with zero technical skills using one move: forward everything to Shraga. It works beautifully — until Shraga relocates to Silicon Valley and Hefer quits, because the whole system *was* Shraga. Most teams shipping AI features today are Hefer: every request gets forwarded to a big commercial model, and it works right up until latency, cost, privacy, or a deprecation notice takes Shraga away. This talk is about building your own Shraga — a small model fine-tuned into a capable agent for one specific job. We built an agent that reads IT logs, diagnoses failures, and proposes verified fixes. We'll take it from a vanilla small model that fails almost everything to one that holds its own. Each fine-tuning phase gives it a new personality: the obedient agent (SFT), the self-taught agent (STaR), the picky agent (DPO/DMPO), the opportunist (GRPO/RL). Plus the questions that come first: do we even need an agent, and how do you actually evaluate one?
Lead Applied Research Scientist at Specific AI
Applied Research Scientist at Specific AI