RAVEN.Nest
Financial document AI
that never leaves the building.
Sovereign, on-premise processing for Europe's fiduciaries.
1001 AI LABS · BASEL, SWITZERLAND · PRE-SEED 2026
The Problem
Manual entry costs more than a fiduciary can bill
THE COMPLIANCE TRAP
US CLOUD Act exposure: cloud AI routes client tax and payroll data through US-owned infrastructure.
Swiss nLPD: processing personal financial records in multi-tenant clouds raises data-protection and documented-handling obligations.
members (CH)
fiduciary practices
A paper-heavy back office in every practice.
Full TAM / SAM / SOM on the Market slide.
Illustrative model · ~300 docs/mo at CHF 35/h loaded cost · target base ~250,000 paper-heavy SMEs. Bottom-up estimates.
02 / 16Why Now
New rules are pushing financial data off the cloud
Documented handling and data-residency duties, in force with no grace period since 2023.
France now requires every VAT-registered business to receive structured e-invoices; Factur-X (PDF/A-3 + XML) is the default format. Swiss fiduciaries with French or multinational clients must read and archive the XML, not just the PDF.
GoBD demands immutable, documented audit trails. EU AI Act obligations may apply depending on the use case and decision context; document extraction itself is not automatically high-risk.
Regulatory scope varies by jurisdiction; phased rollouts; counsel review required.
03 / 16The Solution
Sovereignty by design
An on-premise AI appliance. The documents never leave the client's building.
-
01
Data never leaves the building
Capture, OCR, extraction and review all run on the client's own node. Document contents stay on-site, so there is no US CLOUD Act vector.
-
02
Local AI, zero cloud calls
On-device inference with Mistral, DeepSeek-R1 and Qwen (GGUF) on 32 GB RAM, no VRAM. No external API, no data egress. Apertus 1.5 is a roadmap target, not the current runtime.
-
03
Human in the loop
Every extraction is reviewed on a local dashboard before it is used. Nothing is booked automatically.
Architectural isolation on client-controlled hardware. Hardware configuration under validation.
04 / 16The Real ROI
From cost sink to advisory engine
Automation hands each fiduciary back the hours they can bill as advisory.
Today
25 h/mo manual entry · −CHF 725/mo loss · 70% of hours trapped
With RAVEN
10 to 30 sec/doc · billed CHF 0.10 / 0.50 / 3.00 per doc by pipeline depth · 22.5 h/mo freed
Reallocate to advisory
At CHF 150/hour → yields +CHF 3,375/mo new billing
Value transfer
At 20 SMEs, unlocks
advisory potential per fiduciary firm
Illustrative model (300 docs/SME/mo · CHF 35/h loaded · CHF 150/h advisory · 20 SMEs/firm). To validate in pilot; not realised revenue.
05 / 16Market
Fiduciaries are the wedge into a wider regulated market
BASEL → UPPER RHINE → DACH → FRANCE → EMEA → UAE (PEPPOL / PINT-AE)
B2B2B · THE CHANNEL
Illustrative usage-based scenario: per node ≈ 1,000 docs/mo × CHF 0.50/doc × 12 ≈ CHF 6,000/yr. Practice counts sourced (BStBK, TREUHAND|SUISSE, KSW); the 20 to 50 multiplier and per-node volume are operating assumptions to validate in pilots. Not committed revenue.
06 / 16How It Works · 1 of 2
The whole pipeline runs on one node in the office
Scan / Ingest
Capture documents from any source.
Local OCR + Extraction
Extract text and data entirely on-premise.
Human Review
Validate accuracy and resolve exceptions.
Structured Archive / ERP-ready
Store securely and integrate with your systems.
Current tested mode keeps processing local on client-controlled hardware; ERP integration is a roadmap item.
07 / 16How It Works · 2 of 2
Nine layers turn paper into reviewable books
VALIDATION
Current runtime: Tesseract + local OCR · Mistral + DeepSeek-R1 + Qwen (GGUF) · 32 GB RAM, no VRAM, zero external API calls · human-in-the-loop before any booking. Apertus 1.5: roadmap target (HPE + NVIDIA hackathon, 14 Sept 2026).
08 / 16The Killer Use Case
We clear the quarterly shoebox at its source
In DACH fiduciary practice, “the shoebox” is the bag of mixed paper (receipts, invoices, bank printouts, handwritten notes) a client drops at their Treuhänder each quarter.
Disorganized, multi-format, multi-language, often incomplete. Today the accountant sorts, transcribes and reconciles each document by hand, one at a time.
THE COST · PER TREUHAND FIRM, MONTHLY
- Ingests every format at once, no manual sorting
- Local pipeline extracts VAT, amount, counterparty, date, category
- Human-reviewed output in 10 to 30 sec per document
- Pushes structured data to Bexio / Abacus / DeepBox
- Audit trail stored locally, nLPD compliant, no cloud
- Scan a receipt on the phone
- On-device pipeline: OCR, vision, validation, Swiss audit
- Outputs Factur-X / QR-Bill / ZUGFeRD for the accountant
- Sends straight to the Treuhänder's RAVEN node
Cost ranges are illustrative fiduciary-practice estimates. the on-premise RAVEN node is the MVP today; Corvue is an early-stage idea, built once a pilot proves the need.
09 / 16Validation · Beta Sandbox
A local agent fetches and reads real receipts
Raven MCP Scout, a local MCP agent, logs into the supplier and downloads receipts itself with Playwright, then RAVEN processes them. Local LLM only, zero cloud. Beta-tested on Coop.
WHAT WE SCANNED
INDUSTRIES
Gastronomy · retail · trades & industry · health & pharma distribution
FILE TYPES
Handwritten receipts · multi-page invoices · bank statements · expense notes
WHAT WORKS TODAY
- Reads mixed document formats
- Extracts amounts and entities
- Identifies Swiss VAT
- Runs fully offline, on-device
- Human review before export
SELECTED & REGISTERED
Pre-revenue, pre-seed. Figures come from three offline sandbox runs (provenance in appendix). Still to validate: accuracy on live client data · reconciliation at scale · ERP connectors · recurring willingness to pay.
10 / 16Business Model
Own the node, then pay per finished document
Agent-as-a-Service. One setup fee covers the on-premise node, which the client owns; after that they pay only for finished, booked-ready output.
CapEx-neutral on day one. Nothing leased.
The floor tier: capture only.
Priced by complexity and pipeline depth.
Setup recovers the hardware; usage pricing captures value per document, at a low incremental inference cost after local deployment.
Hardware is optional: buy the node outright or take a managed lease. Enterprise support quoted separately.
Pilot-stage pricing, validated with the first fiduciary partners; "from" rates, not locked. A managed-hardware/support subscription and a full-local-ledger concept are appendix roadmap options.
11 / 16Competition & Defensibility
On-premise plus fiduciary depth: an empty quadrant
▲ On-premise, client-owned hardware · Fiduciary-specialized ▶
Nutanix · Dell AI PCs · local LLM kits
Local, but general-purpose. No fiduciary pipeline or appliance.
On-site appliance + DACH fiduciary pipeline: a combination incumbents don't currently offer.
Findea.ch · Klippa · Mindee · WellyBox
Client data transits US-owned infra: CLOUD Act exposure.
Bexio AI · Abacus
Strong ledgers, cloud-first, no on-site AI extraction layer.
Team & Ecosystem
A founder built for this exact problem
Ten years inside regulated back-offices. The exact stack RAVEN needs, in one founder:
Solo-built RAVEN end to end · 513K+ LOC · Python · FastAPI · PostgreSQL · fully local
COUNSEL & ECOSYSTEM
Roadmap hire: part-time Senior Systems Engineer for node hardening. Anchors are counsel/incubator/mentors, not backers. Einzelfirma; AG planned.
13 / 16The Ask
Four ways to help build the nest
SIVentures anchor · CHF 250K total round (balance from sector angels)
Swiss CLA · CHF 2M cap · 20% discount · 2% p.a. · 24 months
18-month runway. First 3 fiduciary pilots · AG incorporation · senior engineer.
Use of funds: 35% engineering & deployment · 25% pilot hardware · 20% AG + IP filings · 20% Basel customer development.
Fiduciary or accounting firm. Run RAVEN on your documents for 60 days, no commitment, full sovereignty. Named as a founding design partner.
→ alaasalathe@1001ailabs.com
Receive a RAVEN Scout node on-site. 60-day pilot with your real document corpus; we handle setup, support and reporting.
→ 3 slots · Basel region priority
A finance or AI researcher (PhD, postdoc, or university student) to co-develop and benchmark local document AI on Swiss data. Innosuisse / academic fellowship route.
→ university collaboration welcome
18-MONTH MILESTONE MAP (targets)
First design partner signed · benchmark locked
2 to 3 pilots live · first nodes deployed on-site
5 to 10 fiduciary clients · nodes across the Basel region
Seed round · DACH expansion (Freiburg, Mulhouse)
20+ nodes deployed · founding team in place
Targets, not guarantees. Swiss AG planned; pre-revenue. Indicative terms subject to due diligence. alaasalathe@1001ailabs.com
14 / 16Join the Journey
What we're building, and who we need to build it
WHAT WE'RE BUILDING NOW
Hardening the on-premise node and its agentic pipeline. We are coding the next build at the HPE & NVIDIA Agentic AI Hackathon, on NVIDIA hardware.
A local MCP agent that logs into a supplier and downloads receipts itself with Playwright, then hands them to RAVEN. Local LLM only. Beta-tested on Coop.
Mobile scan app for SMBs. An early-stage idea, built once a fiduciary pilot proves the need.
THE FIRST TEAM · FOUNDING HIRES
Three ground-floor roles, meaningful equity, shaping the company from day one.
Corvue is an early-stage idea, not a shipped product.
15 / 16Appendix
Every claim traces back to a document
Three-trial register: A 632/476 · B 427/1,534 · C 128/391 → 1,187 / 2,401, Level-A/B grading.
DSO2026013123 (V1.0) + DSO2026013134 (V1.1), ZertES-signed.
513,271 LOC / 892 files, root SHA-256, sole-ownership declaration.
Local MCP agent auto-downloads supplier receipts with Playwright (beta on Coop); RAVEN processes them. Protocol + PDF/A-3 outputs.
Setup fees + hardware BOM (Scout 32GB to Enterprise 192GB, preliminary). Alternatives to the pay-per-output lead model: flat monthly plans (capped at a document allowance) and the full-local-ledger concept.
Local-only isolation architecture.
Data room being assembled now. INPI e-Soleau + ZertES establish date-certain proof of authorship and anteriority.
16 / 16