Document Intelligence · research & products

Document AI that reads — and remembers.

I build systems that understand visually rich documents, ground every answer in evidence on the page, and carry what they read forward as persistent, provenance-linked state, so the amendment that arrives tomorrow is read in the light of the contract signed today.

Portrait of Souhail Bakkali
Souhail Bakkali
Research scientist in Document Intelligence · SHADOC team, IRISA (UMR 6074), Université de Rennes
Now
IRISA · ESIR · Université de Rennes
Before
Associate Professor, L3i, La Rochelle Université (2023–2025)
PhD
La Rochelle Université, 2022 (European Doctorate label)
Company
Co-founder, SynaptYon SARL
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. Outcomes follow the paper’s five reconciliation outcomes; support, retain conflict and escalate for review are named there without a definition, so their use here is our reading.
29
publications
22 peer-reviewed
285
citations
Google Scholar · September 2026
9
h-index
Google Scholar · September 2026
€501.5K
competitive funding as PI / Co-PI
800+
hours taught

§ 01 The problem

Today’s document AI has amnesia, by design.

Models now read contracts, extract fields and point to the page region behind each answer — then forget all of it. My 2026 survey with Ayoub Merimi names this the statelessness bottleneck and specifies what it takes to remove it.

Storage
gives persistence
Files and indexes outlive the session, but nothing is learned from them.
Retrieval
gives access
The right page can be found again, but each document is still read cold.
Consolidation
gives experience
Observations become provenance-linked knowledge that improves the next reading.

“Storage is not retrieval, and retrieval is not consolidation.”

Amnesia by Design, Memory by Necessity · arXiv:2609.32041

Open the interactive explorer of the paper

Persistent state · Eq. 4From the paper

St = (Et, Kt, Γt, 𝒱t, Pt, At)

Et
Episodic evidence buffer
Admitted observations, appended with a timestamp and their provenance.
Kt
Consolidated semantic store
Schemas and facts abstracted from episodic traces.
Γt
Associative relation graph
Relations between entities, documents and schemas.
𝒱t
Temporal validity index
When each claim holds; closed when newer evidence supersedes it.
Pt
Provenance ledger
Links every claim back to the document, region and version it came from.
At
Authorization and policy state
Who may read or change what; updated only by policy events.
Six components of persistent, evidence-grounded document state. Provenance is an invariant across every operation.

§ 02 From papers to products

One stack, from the first read to proof that memory helps.

The memory layer of the SynaptYon chain, as the 2026 survey specifies it. Each layer implements a part of the paper, builds on published work, and maps to one of the memory products. Pick a layer to see the research behind it.

The stack · 7 layers, 6 products Interactive
  1. Acquire Read the whole document

    Extract the multimodal observation: text, layout coordinates, table structure and visual elements, for questions not yet asked.

    Paper
    Observe · Dt = (V, T, B, G, H, M)
    Product
    Mnemos In development
    Research
  2. Validate Gate before writing

    Corroborate each observation against the provenance ledger and reject it when confidence is below threshold, leaving state unchanged.

    Paper
    Pre-write gate · Theorem III.4
    Product
    Argus In development
    Research
  3. Consolidate Write and abstract

    Append admitted evidence with its timestamp and provenance, then abstract recurring families into reusable schemas.

    Paper
    Et, Kt, Γt · max J(α) − λR(α)
    Product
    Mnemos In development
    Research
  4. Reconcile Know what is in force

    Resolve conflicts with one of five outcomes and close superseded claims at their effective time, keeping history for audit.

    Paper
    Temporal supersession · 𝒱t
    Product
    Chronos In development
    Research
  5. Trace Cite the box

    Keep the chain from every conclusion back to document, region and version, through consolidation, supersession and repair.

    Paper
    Provenance invariant · Pt
    Product
    Ariadne In development
    Research
  6. Forget Erase, and prove it

    Remove a target across every store, repair the schemas that depended on it, and certify what can no longer be recovered.

    Paper
    Five-level unlearning · Algorithm 2
    Product
    Lethe In development
    Research
  7. Measure Show that memory helps

    Compare warm, retrieval-only and cold arms on the same future documents and report benefit, cost, risk and governability together.

    Paper
    Longitudinal harness · five metrics
    Product
    Anamnesis In development
    Research
Each layer lists the concept it implements in the papers, the product it belongs to with that product’s current status, and the published work it builds on. Pick a layer; a product chip marks the layers that product covers.

§ 03 Products

The whole chain, from ingestion to answers.

SynaptYon SARL, which I co-founded, builds Document Intelligence for business, financial, historical and healthcare documents: four products that read, extract, answer and remember. None is generally available yet.

  1. Grammata

    Read. Every page, read

  2. Kleis

    Extract. The data you need, from any document

  3. Sibyl

    Ask. Answers you can check

  4. Mnemos

    Remember. Better with every document

Mnemos is the memory layer specified in my 2026 survey; the products page shows them in research detail, with demos.

§ 04 Research trajectory

From fusing text and image to remembering documents.

Since 2019: cross-modal document understanding, then retrieval and grounded answers, low-resource scripts, and now persistent state. Filter by theme.

Research trajectory · 29 publications, 2019–2026 Interactive

2019 1 publication

  1. Face Detection in Camera Captured Images of Identity Documents Under Challenging Conditions ICDAR Workshops 2019 Trust
  2. PhD researcher, L3i PhD Researcher and Lecturer, L3i-Lab, La Rochelle Université

2020 2 publications

  1. Visual and Textual Deep Feature Fusion for Document Image Classification CVPR Workshops 2020 Understanding
  2. Cross-Modal Deep Networks for Document Image Classification ICIP 2020 Understanding

2021 1 publication

2022 1 publication

  1. PhD Ph.D. in Computer Science and Applications, La Rochelle Université, with the European Doctorate Label
  2. Multimodal Document Understanding with Unified Vision and Language Cross-Modal Learning PhD thesis 2022 Understanding

2023 2 publications

  1. Postdoc, L3i Postdoctoral Researcher and Lecturer, L3i-Lab, La Rochelle Université
  2. TransferDoc: A Self-Supervised Transferable Document Representation Learning Model Unifying Vision and Language SSRN 2023 Understanding
  3. Associate Professor, L3i Associate Professor (Maître de Conférences), L3i-Lab, La Rochelle Université

2024 4 publications

  1. State-of-the-Art Khmer Text Recognition Using Deep Learning Models ACET 2024 Text
  2. LLMChain: Blockchain-Based Reputation System for Sharing and Evaluating Large Language Models COMPSAC 2024 Trust

2025 10 publications

  1. DocSum: Domain-Adaptive Pre-training for Document Abstractive Summarization WACV Workshops 2025 Understanding
  2. IDTrust: Deep Identity Document Quality Detection with Bandpass Filtering WACV Workshops 2025 Trust
  3. Fusion of GNN and GBDT Models for Graph and Node Classification GbRPR 2025 Graphs
  4. Evaluating the Impact of Khmer Font Types on Text Recognition arXiv 2025 Text
  5. Confidence-Based Knowledge Distillation to Reduce Training Costs and Carbon Footprint for Low-Resource Neural Machine Translation Applied Sciences 2025 Language
  6. IRISA, Université de Rennes Enseignant-Chercheur Contractuel, IRISA-Lab, ESIR, Université de Rennes (SHADOC team)
  7. Visual Text Generation in Khmer Language: Challenges and Trends with Diffusion Models ICDAR Workshops 2025 Text
  8. Cross-Lingual Learning for Low-Resource Khmer Scene Text Detection and Recognition ICDAR Workshops 2025 Text

2026 8 publications

  1. Cross-Lingual Transfer for Complex Scripts: A Benchmark on End-to-End Khmer Scene Text Spotting WACV Workshops 2026 Text
  2. Investigating Model Compression for Neural Machine Translation in the Biomedical Domain AICS 2025 Language
  3. Terminology-Aware Retrieval-Augmented Knowledge Distillation for Biomedical Neural Machine Translation EAMT 2026 Language
  4. HSB-GDL: A Heterogeneous Subgraph Benchmark for Classification and Query Spotting IJDAR 2026 Graphs
  5. BHCP-DOC: A Real-World Bilingual Dataset for Handwritten Clinical Prescription Detection and Recognition ICPR Workshops 2026 Text
  6. An Empirical Study of Recent Large Language Models for Knowledge Graph to Text Generation on Standard Benchmarks arXiv 2026 Language
Every publication, by year, from the publication list; career milestones from the CV. In the timeline each publication is a dot in the lane of each of its themes (filled: featured work); dashed rules mark the milestones.

§ 05 Selected publications

Selected work.

All 29 publications
  1. arXiv 2026Preprint

    Fresh at Decision, Stale at Action: Routing New Information to LLM Decisions Already Made

    A. Merimi, S. Bakkali, A. Sakat

    A timestamp theory of LLM decision validity and a decision ledger that realises it. Six timestamps define a decision, a freshness ceiling bounds what up-to-date evidence can guarantee at action time, and accuracy factorises into routing new information to affected decisions and judging it for the time of action. Controlled decision streams confirm four predictions across five LLMs.

    arXiv Memory
  2. arXiv 2026Preprint

    Amnesia by Design, Memory By Necessity: Persistent State for Document Intelligence

    S. Bakkali, A. Merimi

    A survey that names the statelessness bottleneck of Document AI, formalises persistent evidence-grounded document state, audits ten benchmarks against eight statefulness criteria, and derives a longitudinal harness with five counterfactual metrics: Experience Gain, Cost Efficiency, Memory Harm, Forgetting Fidelity and Coverage Retention.

    Interactive explorerarXiv MemoryUnderstanding
  3. arXiv 2025Preprint

    MTRACE: Multilingual Retrieval-Augmented Generation for Temporally Diverse Text Corpora

    S. Bakkali, A. Mudet

    A modular multilingual RAG pipeline evaluated on the French and English subsets of MIRACL: semantic query expansion with reciprocal rank fusion, a generation prompt that enforces strict grounding and explicit abstention, and ablations on entity extraction and retriever choice. It gives faithful answers for well-supported queries and abstains correctly on unanswerable ones.

    arXiv RetrievalLanguage
  4. ICDAR 2025Conference

    WildKhmerST: A Comprehensive Dataset and Benchmark for Khmer Scene Text Detection and Recognition in the Wild

    V. Nom, S. Keo, S. Bakkali, M. M. Luqman, M. Coustaty, M. Rusiñol, J.-M. Ogier

    A large-scale dataset of Khmer scene text captured in real-world environments, with 29,601 annotated text lines from 10,000 images, and a benchmark for detection and recognition.

    PDFDOIHAL Text
  5. WACV 2025Conference

    GlobalDoc: A Cross-Modal Vision-Language Framework for Real-World Document Image Retrieval and Classification

    S. Bakkali, S. Biswas, Z. Ming, M. Coustaty, M. Rusiñol, O. Ramos Terrades, J. Lladós

    A cross-modal transformer pre-trained with three self-supervised pretext tasks that unify language and vision, reducing reliance on OCR positional information. It introduces two industrial-style tasks: few-shot document image classification and content-based document image retrieval.

    PDFarXivDOI UnderstandingRetrieval
  6. ACCV 2024Conference

    KhmerST: A Low-Resource Khmer Scene Text Detection and Recognition Benchmark

    V. Nom, S. Bakkali, M. M. Luqman, M. Coustaty, J.-M. Ogier

    A benchmark for detecting and recognising scene text in Khmer, a low-resource script with complex character structure.

§ 06 Work together

Research collaboration, or a product pilot.

Academic

Co-supervision, joint projects, reviewing

I work with groups across Europe and Asia on document understanding, low-resource scripts and grounded retrieval, and co-supervise PhD students.

SynaptYon SARL

Document Intelligence for your documents

If your team handles business, financial, historical or healthcare documents, from invoices and contracts to registers and clinical reports, we would like to hear about it. None of the products is generally available yet; early pilots shape them. Write to sbakkali.sirajai@gmail.com.