Our CTO, Ameya Kanitkar, spent 24 hours testing a new kind of model called Jev and came back a little giddy. On one of our early eval sets, it hit close to 99% accuracy at roughly 1/100th the cost of the LLM-based approach we'd been using.
He explained his thinking on LinkedIn:
Ameya Kanitkar, co-founder and CTO of Larridin, on LinkedIn:
Jev is insane 😱
We have been testing it for the last 24 hours. On one of our early eval sets, it reached close to 99% accuracy at roughly 1/100th the cost of our current LLM-based approach.
Jev is a new “System One” model. It does not generate text like an LLM. It returns classifications, scores and probabilities.
One use case we are testing is measuring AI fluency.
At Larridin, we evaluate agent sessions across dimensions such as context management, prompt clarity, multi-turn thinking, agent steering, thought-partner use and advanced connector use.
We currently run much of this through LLMs, even though the output is ultimately a set of classifications and scores.
With an LLM, each dimension either becomes a separate call, or everything gets packed into one large prompt. Separate calls are slow and expensive. Large prompts can produce inconsistent scoring across dimensions.
Jev can evaluate all these questions in parallel against the same session. Adding more questions does not add proportional latency.
This lets us use a much richer fluency rubric without multiplying the cost or response time.
Twenty-four hours is nowhere near enough testing to draw broad conclusions. We still need to evaluate it against more data and more edge cases.
But we have already identified several other places in our stack where it might fit.
I suspect there are a lot of expensive LLM calls in production today that should simply be classifiers.
Why This Matters if You Run AI in Production
A lot of AI spend goes to jobs that don't need a model to write anything. Sorting tickets, scoring sessions, flagging risk, tagging content: the output is a label or a number. If you're paying a full generative model to produce that, you may be paying for a lot of capability you never use.
What Makes Jev Different
Ameya describes Jev as a "System One" model. It doesn't generate text at all. It returns classifications, scores and probabilities.
That puts it in an interesting gap. Embeddings are cheap but too shallow for real judgment. A full LLM call has plenty of judgment but costs more and takes longer. Jev aims for the middle: richer judgment than similarity search, without another LLM call's latency and cost.
The Use Case We Tested: Scoring AI Fluency
At Larridin, we score agent sessions on things like context management, prompt clarity, multi-turn thinking, agent steering, thought-partner use, and advanced connector use. Today a lot of that runs through LLMs, even though every answer ends up as a score.
That leaves two not-so-great options:
- A separate LLM call per dimension, which gets slow and expensive fast
- One giant prompt that scores everything at once, which can score inconsistently from one dimension to the next
Jev evaluates all the questions in parallel against the same session, and adding more questions doesn't add proportional latency. That means we could use a much richer AI fluency rubric without multiplying cost or response time.
A Healthy Dose of Caution
Ameya was upfront that 24 hours isn't nearly enough to draw broad conclusions. There's more data and more edge cases to get through.
He also took a hard look at enterprise readiness in a follow-up post. Based on TypeSafe's public documentation at the time, parts of the foundation looked solid: a real DPA, GDPR and SCC coverage, subprocessor controls, breach-notification terms, customer audit rights, and a clear statement that customer inputs aren't used for training.
He also flagged gaps a large enterprise would care about:
- Zero Data Retention is available to enterprise customers but isn't the default
- The standard terms don't set a fixed retention period for API inputs
- He couldn't publicly verify SOC 2 Type II, ISO 27001, HIPAA, or similar certifications
- There wasn't enough detail yet on inference telemetry, subprocessors, and data residency
At the time, his take: very excited about the model and the architecture behind it, but for sensitive enterprise workloads he'd wait until that documentation gets clearer.
Update (November 2026): TypeSafe now lists SOC 2 Type II in its Trust Center, along with its subprocessors. Its data-handling guidance also explains telemetry and data processing locations, and says zero data retention is free for all organizations. However, it still doesn't publish a fixed retention period for standard usage.
Closing Thought
Ameya suspects there are a lot of expensive LLM calls in production right now that should simply be classifiers. That's worth checking in your own stack. List every LLM call whose output is a label, a score or a yes/no, and ask whether it needs a generative model at all.
And when a promising new vendor shows up, check retention, certifications, and telemetry before anything sensitive goes through it.
Ameya shares this kind of hands-on, numbers-first stuff on LinkedIn pretty regularly. If you want more of it in your feed, follow Ameya Kanitkar on LinkedIn. He's co-founder and CTO of Larridin.
Related Reading
On larridin.com:
- How to Measure AI Coding Tool ROI for Engineering Leaders: netting recovered developer time against rework and full usage-based spend.
- The Coffee-Break Tax: How One Claude Code Setting Cut Our Token Bill 10–15%: another place AI spend leaks without anyone noticing.