For technical reviewers, investors and infrastructure partners
Technology and evidence
How RAVEN is built, what has actually been tested and what is still a plan. Written for technical reviewers, investors and infrastructure partners.
How to read the status labels
- Tested
- It ran, and a dated output exists that we can show.
- Validated
- Tested, and also checked against something outside our own code: a standard, an external validator or a locked corpus.
- Prototype
- Code exists and runs, but there is no saved result measured against a standard.
- Roadmap
- Designed or planned. Nothing runs yet, or there is no hardware to run it on.
Most of RAVEN is tested. Very little is validated yet, because the locked benchmark does not exist. Building it is the next milestone.
The receipt workflow
The workflow behind Scout. Four of its five stages need no network at all.
Read
Local OCR and line-item parsing, German and French
No networkOn the machine
TestedCheck
Swiss VAT arithmetic; mismatches go to review
No networkOn the machine
TestedArchive
PDF/A-3 with embedded structured invoice data
No networkOn the machine
TestedPrepare
Draft entry for the client’s accounting system
Network: client’s provider onlyClient’s provider
In progressApprove
A person reviews and posts. Never automatic.
No networkClient’s own tool
Tested
Component status
| Component | Status | What it means today |
|---|---|---|
| Local document extraction | Tested | Run on real Swiss retail receipts in two dated controlled tests. |
| VAT arithmetic and review routing | Tested | Routes anything that does not reconcile to a person. |
| Archival PDF/A-3 output | Tested | Conformance markers verified on test files. External validator run is next. |
| Human approval before any ledger entry | Tested | Nothing writes to a ledger without a person. |
| Hash-chained audit log | Tested | Chain integrity verified on recorded test events. |
| Draft handoff to an accounting system | In progress | Not live. Described as live only once the evidence is published. |
| Document classification and retrieval | Prototype | Runs on development machines; not yet measured. |
| Account assignment | Prototype | Maps entries to a Swiss SME chart of accounts. |
| Specialised agents with one approval gate | Roadmap | Parallel agents for reading, matching, VAT and reconciliation. |
| Sensitivity check for governed cloud research | Roadmap | Keeps client data local and lets only non-sensitive research use approved cloud services. |
| Locked benchmark with published protocol | Roadmap | Turns tested results into validated ones. |
Where the architecture is going
Today one pipeline does everything in sequence. The next step is a set of narrow agents that each do one job, all reporting into a single approval gate.
Reading
Layout-aware extraction of text, tables and line items.
Matching
Recognises suppliers and learns from previous runs.
Checking
VAT rules and plausibility checks. Flags and proposes, never overwrites.
Reconciling
Matches bank statements to open items and surfaces anything unusual for a person.
One approval gate
A person approves before anything reaches a ledger.
Deployment paths
RAVEN is designed to run where the documents are. Hardware is a candidate chosen with each client, not a fixed product, and no hardware has been benchmarked by us yet.
RAVEN Scout
Defined by the workflow
One receipt workflow. Deliberately no hardware specification until a unit has been tested.
Under evaluationRAVEN Nest
A single local node
Candidate paths: a unified-memory AMD Ryzen AI Max-class mini PC or an NVIDIA RTX workstation. Large models held in local memory, with no cloud dependency for extraction.
Under evaluationRAVEN Flock
Higher capacity
A DGX-class direction for heavier or multi-entity workloads, once the agents are orchestrated and demand is shown.
Planned
Data residency
- Extraction runs without a network connection. Handing a draft to an accounting system needs one, to the client’s own provider.
- Any cloud stage used for development or validation runs in a Swiss region (Zurich) by preference. If a non-Swiss region is ever used, it is named as such.
- Evaluations use synthetic or authorised data only. Nothing is uploaded through this website.
Next milestones
External validation of the archival output
Run the PDF/A-3 files through independent conformance validators.
Locked corpus and reproducible benchmark
A fixed test set, a published protocol and repeatable results.
First evaluation node on real silicon
Measure the Nest targets on hardware instead of modelling them.
Agent coordinator
The piece that turns the specialised agents into one system.
Infrastructure that moves this forward
For cloud, compute and credit programmes: the workloads we are ready to run, all on synthetic or authorised data.
- GPU inference for benchmarking
- Running candidate models against the locked corpus, in a Swiss region where available.
- Reproducible test runs
- Continuous integration that re-runs the benchmark on every change and keeps the results.
- Model serving for evaluations
- Short-lived environments for client evaluations, torn down afterwards.
Built with startup-programme support from AWS and Anthropic, and GPU access through NVIDIA LaunchPad. No partnership or endorsement is implied.
What we do not claim
- No certification or compliance claims. Architected for Swiss data residency; assessment not completed.
- No performance figures until they are measured on real hardware with a published method.
- No live integrations, no client names and no partner endorsements.
Want the detail behind a label?
We walk technical reviewers through the evidence for any line on this page, and say plainly where there is none yet.