Skip to content
HUIAHUB

Models & inference

ej

An 11 MB open-weight model from HITL Systems that answers routing, yes/no and priority questions about text on CPU.

Verified Open source
Visit official site (opens in a new tab)

What ej does

ej is a small decision model from HITL Systems. It answers fixed questions about a piece of text, such as which team should handle a ticket, whether a tool call is allowed, or how urgent a message is. Instead of free text, it returns a probability for every option, and the probabilities for each question sum to 1.

You install it with pip install ejai and load one file of 11,384,312 bytes that holds the encoder, the heads and the tokenizer. It runs on CPU, works offline and fetches nothing else at load time. A call passes a state (free text or JSON) and one or more questions of three types: choice, noul for yes or no, and score for an ordered level. The options are plain text written at call time, so a new queue or label needs no retraining, and every question about a record is answered in one pass. A method called model.adapt can learn a workflow's label mix from a handful of labelled records without changing the weights.

The vendor suggests labelling a few dozen records from one workflow, finding a confidence threshold with an acceptable error rate, and automating only above it. Its own benchmarks report .721 accuracy on workflows ej was trained on and .419 across 105 workflows it has never seen, where the hosted and larger alternatives it compares against score far higher.

ej suits teams that make the same English-language decisions over and over and need them to run inside their own service. The code is Apache-2.0 and the weights are CC BY-SA 4.0.

  • Routing support tickets to the right team with a probability per option
  • Yes/no checks on tool calls before an agent runs them
  • Scoring the urgency or priority of incoming text
  • Automating confident decisions and sending the rest to a person or a larger model

Who should use it

  • Teams making repeated English-language routing, triage or priority decisions
  • Services that must run on CPU, on their own servers or offline

Who should look elsewhere

  • Brand-new tasks with no labelled data, where the vendor says Jev, Laya or a prompted LLM is more accurate
  • Languages other than English, or running on a phone
  • Decisions about a person's credit, job, health or legal case without human review

Strengths

  • One 11.4 MB file that runs on CPU and works offline
  • Returns a probability for every option, so a threshold can decide what to automate
  • Labels are plain text set at call time, so they change without retraining
  • Apache-2.0 code and open weights

Limitations

  • Accuracy on workflows it never trained on is .419 across 105 workflows, by the vendor's own figures
  • English only
  • Runs as Python and PyTorch on a CPU today, with peak memory of about 558 MB
  • Weights are CC BY-SA 4.0, so changed weights that are shared must use the same license

Sources