SporeLabs

Big models are for demos. Small models are for products.

Your product asks a giant general-purpose model to do the same few jobs, thousands of times a day. We train a model for exactly those jobs. It answers in milliseconds, costs a fraction as much, and runs on our API, in your cloud or inside your app.

Goes straight to Coleman, the founder. Or email coleman@sporelabs.dev.

Four jobs we built models for.

Conversations

Finding commitments, open questions and hand-offs in a meeting

28× faster than Jev, running in your app

  1. Aany last things before we break?Open question5 msOpen question137 ms
  2. AAlright. Fair enough.No cue4 msNo cue145 ms
  3. DWe’re good?Open question4 msOpen question192 ms
  4. ASounds good.No cue9 msNo cue136 ms
  5. AI’m gonna save th a copy of this in case you guys need any reminders.Commitment7 msCommitment131 ms
  6. AI’m gonna save a copy of thisCommitment6 msCommitment183 ms

Ranking on 40 public meetings: commitments and hand-offs, Cue 0.916 and Jev 0.867; open questions, Jev 0.931 and Cue 0.904.

Runs in your app or on our API. Need another cue, like cancel intent? Tell us and we’ll train it →

Assistants

An advising assistant picking its next tool

20× faster than Claude Sonnet

“What culinary programs do you offer?”

Right tool over 100 requests: ours 80, Haiku 84, Sonnet 85.

Real answers, replayed at each model’s average speed.

Privacy

Finding personal details (PII) in messages

7.3× faster on a laptop than Jev’s API over the network

Priya NairName here. Send the contract to priya.nair@northwind.ioEmail, or call me on (415) 555‑0142Phone.

Found on the laptop; the message never left it. Try Scrub →

Video

Cropping video for phone screens

5× cheaper than AWS Rekognition

Per hour of video. Rekognition returns face positions; ours returns the finished crop.

How we measured

Conversations. Medians per line, October 7, 2026, on 116 lines from the AMI Meeting Corpus test split (CC BY 4.0), each with up to three earlier turns. SporeLabs Cue (32M parameters, fp32, one CPU thread of an Apple M3 laptop, each line timed 3 times): 6 ms. Jev through its API from the same laptop, one request per line asking the same three questions, timed round trip, sequential, n=116: 166 ms. The replay shows each system’s real answer for each line after the time it took on that line; the lines are 6 in a row from meeting ES2009b, from “any last things before we break?”. Ranking is AUROC on the full AMI test set (33,612 lines, 40 meetings): commitments and hand-offs, 118 lines AMI marks as action items, Cue 0.916, Jev (v1.13) 0.867; open questions, 1,283 lines, Jev 0.931, Cue 0.904. All results.

Assistant. Averages over 100 held-out advising requests, June 4, 2026: SporeLabs model 0.14 s, Claude Haiku 4.5 1.32 s, Claude Sonnet 4.6 2.82 s. The right tool was picked 80 times by the SporeLabs model, 84 by Haiku and 85 by Sonnet. The culinary request is one of the 100; the calls shown are each model’s real answer to it.

Privacy. Medians per message on an Apple M3 laptop. Jev through its API from the same laptop, one request per message asking six yes/no questions (name, address, phone, email, card, secret), timed round trip, sequential, n=50 messages (median 734 characters), October 6, 2026; SporeLabs Scrub (the int8 ONNX build in the published package, one CPU thread) on the same 50 messages that day, each timed 3 times.

Video. Price per hour of video analysed, at list prices, September 2026. SporeLabs: $0.02 a minute ($1.20) plus $0.005 for one crop track. AWS Rekognition Video face detection: $0.10 a minute, from its pricing page.

Two ways to work with us.

Done for you

We build it with you

We train a model for your job at our cost, prove it matches what you run now, and put it in your stack or your app. Yours to announce.

Talk to the founder

Do it yourself

Use the API

Call Scrub and our other models, or point the OpenAI or Anthropic SDK you already use at sporelabs.dev/v1. Frontier models at cost, no markup.

What it looks like on your bill.

Each drop is one change we made. Your code stayed the same.

Illustrative example

Week 0

Illustrative example: a monthly bill falling to 45% of where it started over twelve weeks, in five changes

Bill 45%

Week 0

Starting bill

    Not a customer’s bill. Percentages are shares of the starting monthly bill.

    What does your AI do most?

    I’m Coleman Isner, the founder. Tell me, and I’ll build your first model with you.

    Or email coleman@sporelabs.dev.