Jev Data Analysis Demos

Comparing MiniMaxH3 acceleration methods across six scenes and reaching up to 9× faster generation without visible quality loss

Analyzing public discussion to reveal how people describe and talk about Upstash in an open-source workflow

Classifying chicks as male or female from textual descriptions of their visible characteristics

Comparing typed decisions with ChatGPT and Claude across two real-work A/B tests

Classifying social posts about Madrid Fashion Week to sort event-related content quickly

Comparing Jev with 30+ open-weight models across 35+ benchmarks and 130,000 questions per model

Fine-tuning Qwen3.5-0.8B with LoRA to score choices directly from logits and return probabilities without prose

Training a competitive Jev-like decision model autonomously with a swarm of agents and an H200 development box

Testing retrieval with typed decisions and documenting where the approach works and fails

Extracting 11 structured clinical variables from free-text notes on consumer GPUs

Filtering deterministic subject-predicate-object candidates into a Constitution knowledge graph in 4 seconds and the Odyssey in 45

Racing typed judgments against Claude Opus, Haiku 4.5, and GPT-5.4 Mini in a shared evaluation arena

Benchmarking GLM 5.3 Flash paired with Jev and OpenJev in a Halite 1 implementation

Benchmarking 662 classifications on SMS Spam and Banking77 against Claude Haiku 4.5 with matched confidence outputs

Handling text, image, audio, and video inputs with a multimodal decision model running under 100 ms on one H100

Indexing the full Katagami style library, converting requests into traits, and scoring candidates for fit in about one second

Auditing 300 web pages with nine questions each for $0.02 and returning type, citability, and fix recommendations

Forcing text generation by selecting the next character from 33 options with a separate call for every keypress

Comparing Jev and Opus 4.8 as supervisors that route work inside the same LangChain multi-agent system

Searching more than 6,000 Y Combinator startups by niche, color, image, age, or competitor in under one second

Benchmarking Apple's default on-device model at about 40 classifications per minute with a 93% pass rate

Compressing communication among large agent swarms with typed decisions and entity extraction for a reported 100–10,000× gain

Extracting product features, ideal customer, messaging, pricing strategy, and business logic from a URL into Markdown

Testing whether 90% confidence is calibrated on 385 real bank-support messages rather than measuring speed alone