Loading builds…
Build with System One models
Projects worth building, step by step, and the apps and tools people have built with decision models.
Build with System One models
Projects worth building, step by step, and the apps and tools people have built with decision models.
Build with System One models
Projects worth building, step by step, and the apps and tools people have built with decision models.
114 builds · from the community
Agents, routers, scorers, apps and games, built on models that decide. Every entry links to the models it uses, so you can pull the same one and start from there.
FEATURED
brier is an open-source Python library that turns open LLMs from Hugging Face into calibrated decision models. You give it a piece of text, such as a support ticket or a message, and typed questions about it: - a choice ("which team should handle this?"), - a yes/no ("is this a refund request?"), - a score ("how urgent, 1–5?"). It returns a probability for every possible answer, read straight from the model's next-token logits. It doesn't generate any text and doesn't fine-tune the model. Raw LLM probabilities are biased by option order and badly calibrated, so brier corrects them in levels: - with no labels (L0): reduces option-order bias; - with 50+ labels (L1): calibrates the confidence; - with 60+ labels (L2): adds a small head on the model's hidden states for higher accuracy. On a 20-intent banking benchmark, 300 labels take six model families to 80–86% accuracy, from 16–71% raw. It also includes brier check <model>, which tests whether a model works before you start, and calibrations you can save and reload. It works with both chat and base models.
brier is an open-source Python library that turns open LLMs from Hugging Face into calibrated decision models. You give it a piece of text, such as a support ticket or a message, and typed questions about it: - a choice ("which team should handle this?"), - a yes/no ("is this a refund request?"), - a score ("how urgent, 1–5?"). It returns a probability for every possible answer, read straight from the model's next-token logits. It doesn't generate any text and doesn't fine-tune the model. Raw LLM probabilities are biased by option order and badly calibrated, so brier corrects them in levels: - with no labels (L0): reduces option-order bias; - with 50+ labels (L1): calibrates the confidence; - with 60+ labels (L2): adds a small head on the model's hidden states for higher accuracy. On a 20-intent banking benchmark, 300 labels take six model families to 80–86% accuracy, from 16–71% raw. It also includes brier check <model>, which tests whether a model works before you start, and calibrations you can save and reload. It works with both chat and base models.
Sree's game describes each moment in one sentence and asks one typed question; the enemies act on the returned probabilities, with either model plugged in.
spoof.sh's ComfyUI workflow chains a Qwen prompt enhancer, a Laya router that picks the style and Qwen-Image on one NVIDIA card — about 37 seconds per image.
Jev against GPT-5-mini, Claude Sonnet 5, fine-tuned DistilBERT, BART NLI and Laya on SST-2, AG News and more, reporting accuracy, macro-F1, ECE, latency, throughput and cost.
A checkpoint trained on PolyAI's 77-intent Banking77 dataset for customer-support classification.
Ahmed had the 322M multilingual checkpoint control every paddle move, running locally on Windows.
Twenty real Chinese requests on every checkpoint, measuring accuracy, latency and whether the model's confidence drops when it is wrong.
FrontierMind used the offline model on a MacBook Air for inbox triage, tagging 100 SAP decks and sorting blog posts, at zero cost per call.
VeidtKV ported Laya to gfx1030 hardware, with a Hermes agent doing much of the porting.
Laya's author points to anth.us's comparison as evidence that task-specific fine-tuning closes, and usually reverses, the gap.
A Unity-native client for local Laya servers and compatible endpoints: game state goes in, typed choices and scores come out.
A native Windows demo on CUDA or CPU showing live probabilities, chosen moves, inference timing and how often the planner stepped in.
Dima Nurm ran the 322M Laya model locally on an M3 Air and filed 1,000 synthetic emails into six folders in 28.6 seconds, at no cost.
A local MLX run of Laya against Jev on real support tickets in two languages. Laya answered faster; Jev was far more accurate, 90–100% against 30–80% zero-shot.
A local prototype that reads a draft and updates three signals live — tone as a choice, formality as a 1–5 score and the risk it starts a fight — in 40–110 ms per check.
Abhijay stress-tested Jev, SemIf and Laya on adapted exam questions: Jev scored 83.7%, SemIf 61.6% and Laya 31.2%.
Describe a past Claude Code, Codex or OpenCode session in plain words; Chat Seek searches the local histories and reranks the matches with Laya.
A reproducible zero-shot benchmark on informal Darija reviews written in both Arabic script and Arabizi.
A frozen benchmark of synthetic Feishu scenarios with fixed inputs, prompts and labels. Laya answered in 151 ms against Jev's 253 ms but got far fewer right.
Laya's author recommends fine-tuning for real use cases and shares a 2×T4 Kaggle notebook that any agentic coding tool can adapt.
A Swift package that reads forms in running Mac apps through the Accessibility API and decides what to fill with Laya on the Neural Engine, around 1 ms per decision.
Madhav Sharma kept the agent and 30 labeled tickets fixed and swapped only the model: Jev reached 53% at 422 ms, Laya 10% at 152 ms.
Nandakishor M's project site and write-up on the multilingual decision engine, its checkpoints, benchmarks and router.
A self-hosted console that runs VirusTotal lookups for IPs, domains and file hashes through Jev, Von or Laya and scores the models against each other.
Many of the first builds here were collected by madewithlaya.com. Every entry links to its creator’s own post, repository or site.