New playbook: when your own AI model pays. Read
Atos (public sector, via it services) reported this on 27 Oct 2026 in vendor case (source). It belongs to the learning loop layer.
A proof of value on Red Hat OpenShift AI: GPT-4.1 generated synthetic training data, then small models (Phi-4 Mini 3.8B, Qwen 2.5 1.5B/7B) were fine-tuned for multilingual structured extraction from social media, deployable locally.
Fine-tuned Phi-4 Mini JSON validity rose from 1% to ~99% (GPT-4.1: 96%); on a 34-point rubric the best fine-tuned model (Qwen 2.5 7B, 30.8) edged GPT-4.1 with a detailed prompt (30.5); training needed at most 3 GPUs.
The number covers a proof of value; production use and cost are not reported.
Reviewed corrections are kept as test cases and later used to train models.
Read the guide to the learning loop.
96%
lower estimated serving cost for one agent
A daily continual-learning loop for Sidekick's GraphQL agent: an expert-calibrated LLM judge scores production conversations, a self-healing pipeline repairs failures, and the repaired trajectories are trained into a smaller model (SFT, then GRPO against the judge) that replaces the frontier model (source).
13.8x
lower cost per task, with higher accuracy than the best frontier model
A Qwen3-235B model fine-tuned on Tinker with investor-labelled data to do six daily information-filtering tasks (e.g. which news matters to a macro investor) (source).
~7x
faster responses from a post-trained open model
A post-trained NVIDIA Nemotron 3.5 Lightning (30B MoE, 3B active) 'task expert' that took over Sidekick's GraphQL sub-agent from a frontier model, served on Shopify's own GPUs (source).