Train a triage model, then decide whether to ship it
Take one agent step from an optimised-prompt baseline to a trained small open model and back to a decision. Measure the baseline and fix the bar, build leak-free training data from an open-weight teacher, fine-tune with QLoRA, train with GRPO in a multi-turn tool environment against a priced reward, hunt the hack you plant, and compare paired, with pass^k, intervals, negative flips and cost per completed case, before writing a ship or no-ship memo.
Quillmoor's triage step reads every incoming ticket, about 20,000 a month, looks the account up and searches the help centre through two stub tools, and returns a category, a priority, the account id, whether a person must take over, a one-line reason and a reply outline. Today Claude Haiku 5.5 does it with an optimised prompt. This project asks the question the whole course has been building towards: would a small open model, trained for this one step, do it better or cheaper, and should you ship it? You'll answer it with numbers you could defend to someone who doesn't want the answer to be yes.
It uses every module. From Decide and prepare: the baseline, the cost model and a bar written before any training, then training data from an open-weight teacher with a leakage audit. From Supervised and preference fine-tuning: a QLoRA adapter on a free GPU, chosen on dev, with optional DPO or KTO for the free text. From Reinforcement learning for agents: a reward priced in dollars and tested before training, GRPO in a multi-turn tool environment, a deliberately gameable reward and the hunt that catches it, regression tiers, and the paired comparison that ends in a memo.
The student is the course's Qwen/Qwen3.5-4B (Apache 2.0, thinking off). Training runs on Kaggle's free T4s, launched and watched from your terminal by Claude Code; the optional paid path is Hugging Face Jobs at $0.40 an hour for a T4 and $0.80 for an L4 as of November 2026, under $10 for the whole project. The teacher and both judges are open-weight models on OpenRouter, about $5 of tokens in all, or free and slower on Ollama for the teacher. The Claude baseline runs through claude -p on your subscription: about 350 short Haiku calls. Expect 8 to 12 hours of your time, plus about 8 GPU hours that run while you do something else.
Quillmoor's triage step reads every incoming ticket, about 20,000 a month, looks the account up and searches the help centre through two stub tools, and returns a category, a priority, the account id, whether a person must take over, a one-line reason and a reply outline. Today Claude Haiku 5.5 does it with an optimised prompt. This project asks the question the whole course has been building towards: would a small open model, trained for this one step, do it better or cheaper, and should you ship it? You'll answer it with numbers you could defend to someone who doesn't want the answer to be yes.
It uses every module. From Decide and prepare: the baseline, the cost model and a bar written before any training, then training data from an open-weight teacher with a leakage audit. From Supervised and preference fine-tuning: a QLoRA adapter on a free GPU, chosen on dev, with optional DPO or KTO for the free text. From Reinforcement learning for agents: a reward priced in dollars and tested before training, GRPO in a multi-turn tool environment, a deliberately gameable reward and the hunt that catches it, regression tiers, and the paired comparison that ends in a memo.
The student is the course's Qwen/Qwen3.5-4B (Apache 2.0, thinking off). Training runs on Kaggle's free T4s, launched and watched from your terminal by Claude Code; the optional paid path is Hugging Face Jobs at $0.40 an hour for a T4 and $0.80 for an L4 as of November 2026, under $10 for the whole project. The teacher and both judges are open-weight models on OpenRouter, about $5 of tokens in all, or free and slower on Ollama for the teacher. The Claude baseline runs through claude -p on your subscription: about 350 short Haiku calls. Expect 8 to 12 hours of your time, plus about 8 GPU hours that run while you do something else.
Free · sign in to start
Sign in to keep reading the brief.
Free, no card. We use your account to save progress, quiz scores, bookmarks and occasionally send updates from our team.