SynaptYon // Document Intelligence

Document AI that remembers.

From the first scan to the last answer. SynaptYon reads, extracts and answers questions on business, financial and historical documents, and remembers what it learned, with a source for every fact.

Four products · one chain · backed by published research

Evidence Templates Links Validity Sources Access
One memory for all your documents: evidence, templates, links, validity, sources and permissions, updated together after every document.

Live demo // one document, every stage

Watch a document go through.

An invoice, a bank statement, a 19th-century register and a handwritten prescription, all fictional. Step through what each product returns, or run it end to end.

Pipeline · one document through the chain Illustrative
Supplier invoice Fictional through the chain
StageProductWhat it outputs
01 IngestGrammata · Capture & quality checks Scanned PDF, 1 page, capture quality 0.96: accepted
02 ReadGrammata · OCR & handwriting Printed French, Latin script; 3 words flagged below 0.90
03 LayoutGrammata · Layout & tables 13 regions in reading order; line-item table rebuilt as 5 rows × 4 columns
04 ClassifyKleis · Classification & routing Supplier invoice (0.97), routed to accounts payable
05 ExtractKleis · Field extraction 13 fields normalised; 1 below 0.90 held for review
06 EntitiesKleis · Names & entities 9 entities, 5 linked to known records
07 VerifyKleis · Consistency & ID checks 5 checks: 4 pass, 1 warning
08 AskSibyl · Questions & answers 3 answers with citations, 1 explicit abstention
09 SummariseSibyl · Summaries 3 sentences, each citing its regions
10 TranslateSibyl · Translation 2 regions, French → English, terminology kept
11 RememberMnemos · Document memory 4 claims written with provenance; matches a known supplier template
Illustrative: scripted outputs on fictional documents. Confidences and scores are examples, not measurements.

Solutions // by document family

Built for the documents you actually have.

Business, financial, historical, healthcare, identity and legal documents: the same chain, assembled for each.

Business // Operations

Business documents

Invoices, purchase orders, delivery notes, contracts, forms, correspondence

  • Classify and split mixed batches the moment they arrive.
  • Extract every field with the box it came from.
  • Catch the amendment that changes last quarter's terms.

UsesGrammata · ReadKleis · ExtractMnemos · Remember

Finance // Audit

Financial documents

Bank statements, annual reports and filings, receipts, audit evidence, tax forms

  • Read tables cell by cell, footnotes included.
  • Trace a consolidated figure back to the receipts behind it.
  • Keep a preliminary statement and its audited version apart.

UsesGrammata · ReadKleis · ExtractMnemos · Remember

Archives // Heritage

Historical documents

Manuscripts, civil and parish registers, printed heritage, low-resource scripts

  • Handwritten and printed text, including scripts most OCR skips.
  • Layouts that were never designed for machines.
  • Names, places and dates linked across thousands of pages.
  • Search a whole collection by meaning, not only by keyword.

UsesGrammata · ReadKleis · ExtractSibyl · Ask

Health // Clinical

Healthcare documents

Prescriptions, lab reports, discharge letters, referrals

  • Handwritten prescriptions, in more than one language.
  • A new report read against the patient's history.
  • Translation that keeps the medical terminology right.
  • Disposal that reaches every record derived from the source.

UsesGrammata · ReadKleis · ExtractSibyl · AskMnemos · Remember

KYC // Onboarding

Identity & onboarding

Identity cards, passports, proof of address, onboarding forms

  • Reject unusable captures before anyone reviews them.
  • Locate the photo and the fields on camera-captured documents.
  • Delete onboarding data when retention ends, and prove it.

UsesGrammata · ReadKleis · ExtractMnemos · Remember

The difference // memory

Read, answer, reset.

Most document AI treats every file as the first one it has ever seen, so an amendment read tomorrow starts from nothing. Read the same four fictional documents both ways.

Live demoExample · stateless vs stateful reading Interactive Illustrative

Contract · MSA-2026-014 · 2026-03-02

MASTER SERVICES AGREEMENT

No. MSA-2026-014 · Rennes, 2 March 2026

1. Parties. This Agreement is made between Halden Freight SAS (the “Supplier”) and Orvelle Analytics SA (the “Client”).

2. Term. Twenty-four (24) months from 2 March 2026.

4.2 Payment terms. Invoices are payable within thirty (30) days of receipt.

5. Rates.
ServiceUnitRate
Standard palletpallet€42.00
Express palletpallet€68.00
Storagepallet-week€3.10

7.1 Liability. The Supplier’s aggregate liability is capped at €250,000 per contract year.

9. Governing law. This Agreement is governed by French law; courts of Rennes.

What a stateful system notes when the amendment and the invoice arrive

  • Supersede Payment terms: 30 days → 45 days from 2026-11-01 (AMD-1 §1)
  • Supersede Express pallet rate: €68.00 → €71.50 from 2026-11-01 (AMD-1 §2)
  • Support Liability cap: unchanged, €250,000 / year stands (AMD-1 §3)
  • Retain conflict Express pallet rate on the invoice: €68.00 vs €71.50 in force (INV-2026-0117)
  • Escalate for review Payment terms on the invoice: 30 days vs 45 days in force (INV-2026-0117)
Fictional documents. The demo runs entirely in your browser; it illustrates the ideas and is not production software.

How Mnemos works // six steps

Six operations. One state.

Every document goes through the same lifecycle. Each step is a product, and provenance runs under all of them.

Source tracing · Runs under every step: any answer traces back to the place on the page it came from. Source tracing →

01 // Acquire

The observation is conditioned on the document, not on a query: state has to hold the full content for questions nobody has asked yet.

  • Text, layout and image togetherCross-modal vision-language reading of the page, in the line of GlobalDoc and VLCDoC.
  • Structure is meaningTable rows and columns, figure–caption links, form–field pairs and section trees are kept as graphs, not flattened to text.
  • Identity and versionMetadata carries document identity, version, timestamp and policy: an amendment means nothing without the version it amends.

You get A complete, structured observation of every document.

Mnemos · Document memory →

02 // Validate

A wrong value written to memory is not one wrong answer. It is a durable error that every later session inherits.

  • CorroborateCheck the observation against what the provenance ledger already supports or contradicts.
  • Score confidenceEpistemic confidence from predictive entropy over the extracted fields; below threshold, the state is left unchanged.
  • Measure the gateEstimate its false-accept rate with controlled corruption, then plan with the contamination bound.

You get Admitted evidence, or an unchanged state.

Mnemos · Quality gate →

03 // Write

Admitted evidence is appended to the episodic store, tagged with when it arrived and where it came from.

  • TimestampedEach entry records its ingestion time, separately from the document's own dates.
  • Source-taggedEach entry links to document, region and version in the provenance ledger.
  • Fast and faithfulThe episodic store is built for quick, high-fidelity writes; abstraction happens later, and slower.

You get An episodic record of everything admitted.

Mnemos · Document memory →

04 // Consolidate

Repeated document families become schemas and relations, chosen by how useful they will be against what they cost to keep.

  • Schemas from familiesTen thousand invoices from one supplier become one template, linked to the evidence it came from.
  • Relations, not just vectorsAn associative graph links entities, documents and schemas by structure, time and analogy.
  • Keep the tailThe optimum is not the smallest representation: episodic detail stays for the rare question the schema cannot answer.

You get Reusable schemas, still linked to their evidence.

Mnemos · Document memory →

05 // Reconcile

New evidence is reconciled against the store into one of five outcomes, never by blind overwriting.

  • Four clocksPublication, event, effective and ingestion time are tracked separately.
  • Five outcomesSupport, supersede, scope-coexist, retain conflict, or escalate for review, and the outcome is recorded.
  • History keptA superseded claim's validity interval is closed, not erased, so audits can see what changed and when.

You get Claims with validity intervals, and their history.

Mnemos · Versions & amendments →

06 // Forget

A deletion request reaches every representation derived from the target, including schemas that absorbed it.

  • Beyond the fileRaw files, page images, OCR caches, embeddings, graph nodes and index entries.
  • Schema repairRe-consolidate what depended on the deleted evidence; this is where current systems stop.
  • CertificateProbe-based bounds on residual leakage, reported with Forgetting Fidelity and Coverage Retention.

You get A repaired state and an audit certificate.

Mnemos · Verified deletion →

Built for production // six commitments

Documents you can act on.

The harness // simulated

What the harness reports.

This is how every pilot is reported: the same model on the same future documents, warm, retrieval-only and cold, with the five metrics side by side.

Mnemos · evaluation on your documents Simulated
  • EG Experience Gain {"id"=>"eg", "abbr"=>"EG", "name"=>"Experience Gain", "equation"=>"EG(k) = Q(D_k+1 | S_k) − Q(D_k+1 | R_k)", "measures"=>"Benefit of consolidation beyond retrieval access.", "good"=>"positive"}Benefit of consolidation beyond retrieval access.
  • CE Cost Efficiency {"id"=>"ce", "abbr"=>"CE", "name"=>"Cost Efficiency", "equation"=>"CE(k) = 1 − C(D_k+1 | S_k) / C(D_k+1 | S_0)", "measures"=>"Inference cost saved by warm state, in tokens or FLOPs.", "good"=>"positive"}Inference cost saved by warm state, in tokens or FLOPs.
  • MH Memory Harm {"id"=>"mh", "abbr"=>"MH", "name"=>"Memory Harm", "equation"=>"MH(k) = Q(D_k+1 | S_k^clean) − Q(D_k+1 | S_k)", "measures"=>"Degradation induced by state; positive values mean memory hurt.", "good"=>"bounded"}Degradation induced by state; positive values mean memory hurt.
  • FF Forgetting Fidelity {"id"=>"ff", "abbr"=>"FF", "name"=>"Forgetting Fidelity", "equation"=>"FF = 1 − Leak(X_del | S′_t) / Leak(X_del | S_t)", "measures"=>"How much of a deleted target still leaks from the post-deletion state (probe-based mutual-information bound).", "good"=>"high"}How much of a deleted target still leaks from the post-deletion state (probe-based mutual-information bound).
  • CR Coverage Retention {"id"=>"cr", "abbr"=>"CR", "name"=>"Coverage Retention", "equation"=>"CR = exp( Q(Y_ret | S′_t) − Q(Y_ret | S_t) )", "measures"=>"Whether knowledge that should survive a deletion still does.", "good"=>"high"}Whether knowledge that should survive a deletion still does.
Simulation with illustrative parameters. These are not measurements of any system.

Work with us // three steps

From your documents to production.

01 // DiscoveryWhere does Document AI pay in your flows?

We map what arrives, in which volumes and formats, who reads it and which decisions follow, and we check how well any existing pipeline carries knowledge from one document to the next.

  1. Document flows mapped: sources, volumes, formats, owners.
  2. The decisions each document feeds, and what an error costs.
  3. The products that matter for you, and the metrics to judge them.

You leave with A scoped plan: which products, on which documents, judged how.

Ask about it →

02 // PilotProof on your own documents.

We run the chain products you need on a stream of your documents, with the accuracy targets you set per field and the harness metrics: warm state against retrieval-only and cold processing.

  1. A stream built from your documents, with session boundaries and the edge cases you care about.
  2. Field-level accuracy on your labels, plus Experience Gain, Cost Efficiency, Memory Harm, Forgetting Fidelity and Coverage Retention.
  3. A review of every error class with your team.

You leave with Numbers on your data, and a go or no-go.

Ask about it →

03 // ProductionRolled out, and owned by you.

We integrate the pipeline with your systems on the infrastructure agreed in the pilot, with validation gates, version history, citations and deletion policies set for your domain.

  1. Integration with your intake, storage and business applications.
  2. Gate, reconciliation, retention and access policies for your documents.
  3. Documentation, runbooks and the harness to keep measuring.

You leave with A document pipeline you own and can audit.

Ask about it →

The names // σύναψις

A joining together.

SynaptYon
σύναψις

synapsis

Greek: a joining together, a conjunction. Also the root of “synapse”, where one neuron hands a signal to the next.

That is what the products do: they join reading, extraction, answers and memory into one chain, so each step builds on the others.

The people // behind the state

Research first.

SynaptYon comes out of Document AI research published since 2019: cross-modal document understanding, grounded retrieval, low-resource scripts, and document memory.

Souhail Bakkali
Souhail Bakkali Co-founder · research & technology Lecturer-researcher at IRISA, Université de Rennes · PhD, La Rochelle Université · Document AI research since 2019 Research site and publications →

Published research

29 publications since 2019, at venues including WACV, ICDAR, Pattern Recognition and IJDAR.

All publications →

Where the research was done
  • IRISA (UMR 6074)
  • Université de Rennes
  • CNRS
  • ESIR
  • L3i, La Rochelle Université
  • ENSA Berrechid, Hassan First University
Published at
  • WACV
  • ICDAR
  • Pattern Recognition
  • IJDAR
  • ICIP
  • ACCV
  • CVPR Workshops
  • ICPR Workshops
  • EAMT

Start with your documents.

Tell us what arrives, in which volumes, which decisions it drives and how often it changes.

sbakkali.sirajai@gmail.com

Opens your email app with the message filled in; nothing is sent from this page.