SoftAI / gut-health engine by Soft All Things

Gut health,
read from a photo.

SoftAI is the engine behind PoopCheck. Send one stool image and get back nine structured readings, a plain-language summary, and a memory of the user that every later call builds on.

Request API access See the API
POST /softAi/stool/analyze example analysis
field reading confidence models
bristolTypeType 4863/3
colorbrown923/3
consistencynormal792/3
shapesausage843/3
quantitynormal712/3
healthhealthy883/3
bloodnone963/3
mucusmoderate precautionary641/3
floatingsink732/3

screenings

  • shape
  • consistency
  • color
  • blood
  • mucus
  • quantity

analysis

Nicely formed Type 4, normal brown color, no sign of blood. There’s some visible mucus, which on its own is usually nothing. If it keeps showing up over the next week, mention it to a doctor.

actionToConsider

Add a glass of water before breakfast for the next few days.

mucus, 1/3  One of three models read “moderate” at 64% confidence; the other two said “trace”. On safety fields, SoftAI reports the highest confident reading instead of letting it get outvoted.

197,313
stool logs from PoopCheck users live
29,489
people who have logged with PoopCheck
139,562
labeled records in our public Gut Health Index
Jul 2024
in production, every day, since

01The engine

What happens between the upload and the JSON.

Five stages, one request. Add ?stream=1 and each stage arrives as a server-sent event while it runs, so your UI never sits on a spinner.

  1. 01 · validate

    Is it actually stool?

    Selfies, lunch and screenshots get turned away before any analysis runs, with a message your app can show as is.

  2. 02 · read

    Three models, independently

    Our own vision model, trained on PoopCheck images, reads the photo alongside two frontier vision models. None of them sees the others’ answers.

  3. 03 · reconcile

    Vote, then be careful

    Majority wins on each field, with a fixed trust order when nobody agrees. On blood and mucus, a confident higher reading escalates the result. It never de-escalates.

  4. 04 · explain

    Words a person can use

    A language model turns the readings into a short summary, a green, yellow or red screening per field, and one specific thing to try.

  5. 05 · remember

    Every call adds context

    With your user ID attached, the result is stored as an event, distilled into memories, and folded into a profile the next call reads from.

02The readings

Nine readings. A closed vocabulary.

Every field comes back from a fixed set of labels with a 0–100 confidence. Store it, chart it and branch on it without parsing a single sentence of prose.

fieldlabelsscale
bristolTypeType 1234567ordered
consistencyhardsoftnormalliquidset
shapesausagelumpyflatblobliquidset
quantitysmallnormallargeordered
colorbrownblackwhitegreenyellowredorangeset
healthhealthyunhealthyset
bloodnonetracemoderatehighordered · safety
mucusnonetracemoderatehighordered · safety
floatingsinkfloatset

a → b ordered scalea · b unordered setsafety escalates, never averages

03Over time

One photo is a reading. A month of them is a pattern.

The same user ID that powers memory unlocks the longitudinal endpoints. They’re what PoopCheck builds its dashboard, reports and assistant from.

  • /stool/gut-score

    A 0–100 score across four dimensions: consistency, regularity, comfort and lifestyle, each with a short and a long explanation.

  • /stool/insights

    Analytics over a window you choose, 30 days by default, written up for the user rather than for a statistician.

  • /stool/health-check

    Stool-type distribution and core gut-health metrics in one call.

  • /stool/lifestyle-tips

    Concrete tips grounded in this user’s history, not a generic list.

  • /stool/chat

    A gut-health assistant that remembers the user across sessions and knows what their last scan showed.

  • /stool/recommendations

    Diet and habit suggestions from recent logs.

04Integration

Built to drop into a product you already run.

Plain HTTPS and JSON, one bearer key, no SDK required. If you can upload a photo, you can call SoftAI.

One key per organization
Keys are stored hashed. Your users, images and memories are isolated from every other tenant.
No user sync
Send your own ID as externalIndividualId. We create the profile the first time we see it.
Real deletion
DELETE /softAi/user/{id} erases images, memories, context and logs in one idempotent call. Built for GDPR erasure requests.
Auditable readings
Every model’s vote is kept with each analysis, so any reading can be traced back to what each model saw.
request 
curl https://api.softallthings.com/softAi/stool/analyze \
  -H "Authorization: Bearer $SOFTAI_KEY" \
  -F "image=@photo.jpg" \
  -F "externalIndividualId=user_8f31"
response200 OK
{
  "primaryPrediction": { "type": "Type 4", "confidence": 86 },
  "secondary": {
    "color": [{ "label": "brown", "confidence": 92 }],
    "blood": [{ "label": "none", "confidence": 96 }],
    "mucus": [{ "label": "moderate", "confidence": 64 }]
    // + consistency, shape, quantity, health, floating
  },
  "screenings": {
    "mucus": {
      "status": "yellow",
      "recap": "Some mucus visible"
    }
    // + shape, consistency, color, blood, quantity
  },
  "analysis": "Nicely formed Type 4, normal brown color…",
  "actionToConsider": "Add a glass of water before breakfast…"
}

05Data & research

Most stool datasets come from a lab. Ours come from bathrooms.

Ordinary phones, ordinary light, real variety, often logged alongside the context that makes an image useful: food, water, sleep, stress and pain. That’s the data a model meets after launch, and it’s very hard to collect in a clinic.

logs
197,313 since July 2024, iOS and Android
labeled
139,562 records in the Gut Health Index, Sept 2026
context
Optional answers on meals, water, sleep, stress, pain and duration
consent
Explicit, before the first scan, including de-identified licensing for research
  1. A

    API access

    Run the engine inside your own app, under your own brand, with the same endpoints PoopCheck uses.

  2. B

    Dataset licensing

    Labeled, de-identified real-world images to train or validate your own models. Scope and terms are agreed per partner.

  3. C

    Research partnerships

    Design a study or a model with us, on top of a community that already logs every day.

Built for teams working on

  • Digital therapeutics
  • IBS & IBD care programs
  • Telehealth & remote monitoring
  • Microbiome & nutrition research
  • Clinical AI validation
  • Food & supplement brands

06In production

We’re our own first customer.

PoopCheck, our app on iOS and Android, runs entirely on SoftAI. Every scan, score, report and chat reply goes through the same endpoints you’d call, so the edge cases reach us before they reach you.

PoopCheck dashboard showing a gut health score of 56 out of 100, broken into consistency, regularity, comfort and lifestyle
/stool/gut-scoreThe dashboard score
PoopCheck digestive health report with a Bristol type distribution chart over six months
/stool/health-checkThe health report
PoopCheck's SOFTai assistant opening a conversation about the user's latest scan
/stool/chatThe SOFTai assistant

poopcheck.app App Store Google Play

07Contact

Talk to the people who built it.

Soft All Things is a small team. You’ll talk to the people who trained the model and run the app, not to a sales funnel. Tell us what you’re building and which part you need: the API, the data, or a study.

marco@softallthings.com