Talk to us
ThesisTrain on how your team actually works →

Models shaped by
how your team works

Capture the sessions your experts already run. Train a small model on that work. Keep the weights. Run it on your own machines.

Talk to us →Read the thesis

The bet

General models learn the public web. Yours should learn your job.

Frontier models are good at a lot of things. They still don’t know how your team closes a ticket, which internal tools matter, or which shortcuts get rejected.

So people paste the same context into prompts. Again and again. Midbar starts from the sessions themselves: what an expert looked at, which tools they called, what they fixed when it went wrong.

From that, we train a small model for one kind of work. Not a thinner generalist. A specialist whose weights carry local procedure.

The model should know how your team does the job, not just how the internet talks about jobs like it.

How it works

Sessions in. Owned model out.

Same loop every time. Capture work, train on it, measure against real jobs, then feed misses back in.

01

Record

Log real sessions: context opened, tools called, edits made, dead ends abandoned, and the outcome that stuck.

02

Clean

Turn messy traces into training examples. Drop noise. Keep the decisions that actually decided the work.

03

Train

Fit a small model to that workload. Score it on held-out jobs of the same kind, including the tool path, not just the final answer.

04

Keep going

Ship it. Collect the cases it mishandles. Retrain. The model should get better as the team’s work drifts.

Compared to what you already have

A specialist is a different product

Frontier model
Generic small model
Midbar workload expert
What it’s for
Many domains, average competence
Same idea, cheaper and weaker
Private context
Pasted into a long prompt, every turn
Same paste, smaller window, more truncation
Training data
Public and licensed corpora
Public data, often plus synthetic text
Where it runs
Provider API
Cloud or laptop, still generic
Cost shape
Pay frontier rates on every turn
Cheaper, still generic
Who owns it
The provider
Provider or open weights

We’re not chasing a smaller GPT. Click a Midbar cell for a short note.

Deployment

Train where the data is. Run where the work is.

Some jobs can’t leave the building. Some can’t wait on a remote API. Midbar is for those: train on private sessions, then run the specialist in your VPC, on a laptop, or on a device.

  • Data stays put

    Session logs don’t leave your environment for someone else’s training cluster.

  • You get a model file

    Not a forever API bill to Midbar. Train, take the weights, run them where you need them.

  • Smaller where it counts

    Once the job is narrow, you can usually run it on far less hardware than a frontier call.

Evaluation

Score the path, not just the final answer

A lucky correct output can hide a bad process. We grade specialists on held-out jobs from your own work, including the tool sequence.

Did it finish the job?

Right answer, right artifact, accepted outcome.

Did it take a sane path?

Right tools, in an order a teammate would recognize.

How often did a human intervene?

Corrections are signal. High rates mean the model isn’t ready.

Would you trust it on a real ticket?

Held-out work from your queue beats a public quiz every time.

  • Expert path
  • Midbar (should match)
  • Generic + human fixes (✕)
One real ticket. Same steps. Midbar is graded on whether it took the expert’s path - not only whether the final answer looked right.

Animation of one held-out ticket. Expert and Midbar complete the same tool steps. A generic model picks wrong tools and needs human corrections marked with X.

Good fits

Work that repeats. Work that’s expensive to re-explain.

Engineering

  • Repo-specific refactors and migrations
  • Debugging that follows your internal playbook
  • Incident digs that look the same every week
  • Test suites that mirror how your team writes them

Operations

  • Triage with your severity rules
  • Runbooks that touch internal systems
  • Escalations that shouldn’t skip a step
  • Reconciliation jobs with house constraints

Judgment-heavy queues

  • Routing and approval gates
  • Classification against your taxonomy
  • Retrieval that prefers your sources
  • Decisions where “close enough” isn’t enough

Notes

A few longer reads

Start with the thesis if you want the full argument. These are shorter cuts into data, eval, and deploy.

Your team already knows how to do the work

Let’s put that into a model you actually own.

Talk to us →Read the thesis