Inbrev
Hack Apertus 2026 · Track 2B · Team Apertool
Run locally, the practice inbox never leaves the practice.
Inbrev sorts the inbox of a Swiss GP practice by handling path and
deadline, with Apertus, built to run on a box in the practice. Every
field it extracts quotes its source.
solo builder · fictional data only
The problem
Before the first patient, there is the inbox.
Lab reports, refills, insurer forms and faxes in four languages land
in one queue. A practice assistant (MPA) reads them one by one.
9.5h / week
of administration for Swiss GPs, up from 8.6 h in 2020
Workforce Study, Praxis 2026 · n = 1,776
54min / day
on requests from authorities and insurers alone
gfs.bern for FMH, 2025 · N = 1,532
Art. 321
Professional secrecy binds doctors and their assistants
Swiss Criminal Code · FADP Art. 5
How it works · Purposeful use of AI
Sorted by who handles it and by when.
01 · READ
Mail, forms, faxes
E-mail, IMAP, web forms, fax via OCR. German, French, Italian, English.
02 · ROUTE
9 handling paths
Plus the administrative deadline and who takes it.
03 · PROVE
Quote or drop
Every field quotes its source. No quote in the original, no field.
Clinical contentPossible symptoms leave automation and go to a person. Not rated, not ranked.
Demo
One fax, sorted. Every value shows where it came from.
1
2
1Route and deadlineInsurer, admin, by 8 Oct. The date comes from the quote, the deadline from code.
2Quote or dropEach highlighted value is verbatim in the fax. No quote, no value.
Running app, stored Apertus 8B result, fictional fax. Possible symptoms would go to a person instead.
Architecture · Sovereign deployability
Pull the cable.
Inbrev keeps sorting.
Inside the practice · no outbound connection
On-premmini PC in the practice
Air-gappedmail gateway hands messages in
Swiss cloudoption for single practices
Docker, no network
egress dns / tcp / https:
FAILED
Local Q4, laptop iGPU
71.8 s / message, 16/16 clinical out
Evaluation · Technical rigour
No synthetic clinical message ran by itself.
180 synthetic messages, labels checked by one person on 60.
| Setup |
Routes right |
Clinical out |
Runs by itself |
| Keyword rules + lexical net |
0.79 |
7/16 |
88 % |
| Apertus 8B, one answer, no checks |
0.87 |
9/16 |
95 % |
| Inbrev with Apertus 70B |
0.96 |
15/16 |
82 % |
| Inbrev 8B + symptom pass |
0.94 |
16/16 |
78 % |
16/16
clinical messages out of automation
95 % CI 0.81–1.00 (Wilson)
- Labels fixed before any text; Apertus never writes or labels test data.
-
Two open models (not people) relabelled 367 messages blind: κ 0.93 route,
0.97–0.98 clinical. One person, blind, 60 messages: κ 0.98 route, 1.00 clinical.
Purposeful use of AI · model choice
Why Apertus, and what it costs.
Measured, same pipeline
- Qwen3.5-9B, Ministral 8B: 177/180 routes
- Apertus 8B: 170/180 · 70B: 172/180
- Fresh hold-out: 5 wrong routes run by themselves (Qwen 0); 70B second opinion: 3
Why Apertus anyway
- Only openly available training data, opt-outs applied, personal data filtered
- Published data catalogue and EU training-content summary
- Local run and quote checks work with any model: no argument
Value, cost & scalability
Swiss francs per 1,000 messages, on hardware the practice owns.
On-prem mini PC
≈ CHF 12
at 60 messages a day; box, power, CPU time before the symptom pass
Swiss cloud VM
CHF 2.39
billed while sorting; same CPU time, other CPU not measured
Swiss token API
CHF 3.13
Apertus 70B list price, tokens measured with the symptom pass
≈ 32 min
MPA time saved a day assumption
60 messages × 78 % automatic × 45 s saved, minus rework of misroutes
Intended purpose · Implementation feasibility
Administrative by default. Clinical mode is research only.
Default mode: Inbrev does
- Sort by path and deadline
- Extract fields, each with its quote
- Say “you decide” when unsure
Default mode: Inbrev never
- Rank by clinical severity
- Rate symptom severity or suggest care
- Send patient data out (local mode)
Business model & roadmap
A module for the software practices already use.
- User
- The MPA, first thing in the morning
- Buyer
- Practice owner or doctors' network
- Price
- CHF 150–250 / month hypothesis
- Channel
- Module in HIN mail or practice software
OCT 2026
Evaluation, offline proof, cost
NOV – DEC 2026
5–10 interviews with MPAs and owners
Q1 2027
Pilot in one group practice
Q2 2027
Classification review before first sale
Inbrev
Team Apertool · Niklas
Wichter, solo builder
Our ask
Pilot Inbrev in one Swiss group practice.
On its own box, from Q1 2027. Your inbox stays in your practice.Contact: hello@apertool.ch