Applied AI on open models
AI that does the job, and shows its work.
Finished tools for real work: check a brief, run the visit, fill the form, make a story. Try one on a sample, call it by API, or run it on your own hardware (on request).
Every result points at its sources, and every hosted model call comes with a signed receipt.
Start here
Pick a job. See it done.
Each one opens on a sample, so you see a real run before you sign up. The numbers come from each tool's own tests.
- HealthcareRecorded run
Patient: …since yesterday my left leg feels kind of weak.
Worth covering: bladder or bowel control problems · NIAMS
Clinician: Any trouble controlling your bladder or bowels?
She denied bladder or bowel incontinence. ↩ turn 23 · checkedRun the visit with a copilot
One screen for the whole visit: who said what, what is still worth asking with its source, a note checked sentence by sentence before you see it, codes and prior-auth criteria, and the work note, FMLA sections and patient instructions drafted, never signed or sent. On your own GPU for real patients.
- 8.4%
- Medical terms missed by the live transcript (lower is better)
- 60 s
- typical (median) on the sample, measured 29 Sep 2026 on our own hardware; slowest 1 in 20: 72 s
- LegalLive
Brief check · sample
problem925 F.3d 1339: no case starts there
problemT.L.O. at 341: “permit” should be “require”
problemPage 1: a full Social Security number
Each finding links to the source it was checked against.
Check a brief before filing
A checklist pointing at the exact words in the brief: fake or misread authorities, misquotations, bad record cites and personal identifiers left in.
- 89%
- Fake citations flagged as a problem (held-out test)
- 12 s
- typical (median) on the sample, measured 29 Sep 2026 on our own hardware; slowest 1 in 20: 19 s
- CreateRecorded run
“I am Pip,” he squeaked. “I cannot see the stars. They are too high, and I am too small.”
page 2 of a Studio storyMake a story, a song or a film
Your own cast and voices, with consent. Every file labelled as AI.
- ~$0.029 per image
- measured 25 Sep 2026 on our own hardware, at list price
- OperationsPreviewRecorded run
Claim form (PDF) · filled from your papers
Policy no.HM-4471-20933policy-letter.pdf:8Date of loss03/14/2026police-report.png:8Repair total1963.18repair-invoice.pdf:15Signature✋ left for you: you check, sign and send itFill a form from your own papers
Your PDF form filled (or a copy list for a form on a website), every answer next to the line of your papers it came from, and the signature and the rest left for you.
- 1 / 491 (0.20%)
- Wrong answers among all answers it wrote (PDF forms)
- 14 s
- typical (median) on the sample, measured 28 Sep 2026 on our own hardware; slowest 1 in 20: 17 s
Build with it
One key, every tool and model, a receipt per call.
The hosted API speaks the OpenAI format, so any SDK or coding agent works. Pay as you go in Decosa credits: $1 each, in US dollars. Self-hosting the tools is on request while it's in early access.
curl https://api.decosa.ai/v1/chat/completions \
-H "Authorization: Bearer $DECOSA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.8-27b",
"messages": [{"role": "user", "content": "Check this clause for risks: ..."}]
}'
# The answer, plus a signed receipt id in the x-decosa-receipt header.Try one now
Three to make, three to check.
Make and do
Check and prove
- Answer a payer audit9 / 9 Weak claims flagged on test audits it had never seen (held-out test)
- Find accessibility fixes for a shop122 / 127 Planted accessibility problems caught (held-out test)
- Check their discovery responses174 / 235 (74%) Deficient responses caught on unseen real cases (motions to compel; held out)
Start to finish
Tools chained on one matter.
Each step hands its result to the next, on the same case, practice or prompt.
- Decosa Legal
Answer their motion
One matter: check their brief, digest the deposition they rely on, then check your reply before filing.
- 01Check the other side's brief
- 02Digest a deposition
- 03Check a brief before filing
- Decosa Clinics
Keep your practice audit-ready
For a therapy practice: notes from your jottings, checked against each client's payer before billing; an audit answered from the notes already on file.
- 01Write a note from my jottings
- 02Answer a payer audit
- Developers
Move a prompt off GPT
Check an open model can take over the prompt, call it through the OpenAI API, and check the provider you'll pay serves the model it claims.
- 01Check if an open model can take over your prompt
- 02Run an open model behind the OpenAI API
- 03Check your provider serves the model you pay for
Trust
You can check our work.
Where it runs
Proof
- 69tools with a held-out testAccuracy, time and cost per run, with the caveats.Accuracy
- Ed25519signed receiptsHashes of what went in and out, signed with Decosa's key. Anyone can check one.Verify a run
- 85tools timed on productionEach sample's median time on production, with the date it was measured.Status
Apps by Decosa
The Decosa family
Separate apps from Decosa, Inc. Each one says what it uses from Decosa.
Meet the family
