You will get DocAI — Offline Document Intelligence Pipeline
Project details
Privacy-first document extraction that runs entirely offline using open-source VLMs and LLMs. Supports invoices, contracts, POs, bank statements, and ID cards across 4 extraction modes (End-to-End VLM, Hybrid OCR+LLM, Graph-Based, Multi-Page VLM). Each mode runs through ingestion → OCR → classification → field extraction → validation → confidence scoring → export. Features RAG retrieval with few-shot examples, arithmetic/format validation rules, anomaly detection, and calibrated confidence scoring. Includes a real-time React/TypeScript UI with WebSocket pipeline control, human review with inline corrections, and an Ace AI assistant. Exports to UBL 2.1 XML, EDI 810, CSV, or JSON. No cloud, no API keys, no data ever leaves your machine.
Machine Learning Tools
Azure Machine Learning, BERT, ChatGPT, Cloudera, fastText, Google AutoML, GPT-3, Keras, Microsoft Excel, MLflow, NLTK, NumPy, OpenCV, pandas, Python, Python Scikit-Learn, PyTorch, scikit-learn, SciPy, SQL, Tableau, TensorFlow, Tesseract OCR, Vertex AIWhat's included
| Service Tiers |
Starter
$350
|
Standard
$1,500
|
Advanced
$4,000
|
|---|---|---|---|
| Delivery Time | 3 days | 7 days | 14 days |
Number of Revisions | 2 | 3 | Unlimited |
Number of Model Variations | 1 | 1 | 1 |
Number of Scenarios | 1 | 1 | 1 |
Number of Graphs/Charts | 1 | 1 | 1 |
Model Validation/Testing | - | ||
Model Documentation | - | ||
Data Source Connectivity | - | ||
Source Code | - | - |
About Bakkali
AI Research Scientist | Ph.D. | Document AI, RAG & Custom ML Solutions
Rennes, France - 2:25 am local time
I hold a Ph.D. in Computer Science (AI & Machine Learning, European Doctorate) specializing in multimodal document understanding. I bridge research and production — designing novel architectures published at WACV, Pattern Recognition, ICDAR, and ICIP — while shipping production-grade systems that process thousands of documents offline with measurable accuracy.
How I work:
✦ 30-min strategy call → I deliver a written report: model recommendations, data plan, architecture diagram, cost vs. accuracy trade-offs, and a realistic timeline. No fluff — just what you need to make a decision.
✦ Full engagement → I own delivery end-to-end: data annotation pipeline, model design/training/evaluation, API/CLI development, Docker deployment, and integration with your ERP or document management system.
✦ Production mindset → Every system ships with validation suites, confidence calibration, audit trails, human-in-the-loop review, and export to your target format — UBL 2.1 XML, EDI 810, CSV, JSON, or custom API.
Book a consultation for straight answers on:
• What's the fastest path to automating your document workflow?
• Open-source fine-tune vs. API — which saves you more long-term?
• How much data do you actually need, and how should you annotate it?
• What accuracy is realistically achievable, and how do we measure it?
• How to keep sensitive documents private (no cloud, no data leaks)?
Specialties:
📄 Document AI & Production-Grade Extraction — I built DocAI, a local-first pipeline that extracts structured fields (invoice numbers, totals, line items, clauses) from PDFs and images using 4 extraction modes — VLM, Hybrid OCR+LLM, Graph-Based, and Multi-Page VLM. WebSocket-driven step control, configurable RAG retrieval, confidence scoring, anomaly detection, and human review with inline corrections. All offline, no API keys. Exports to UBL 2.1 XML, EDI 810, CSV, JSON.
🧠 Multimodal Vision-Language Models — Ph.D. research in cross-modal VL architectures for documents. Published GlobalDoc (WACV 2025) and VLCDoC (Pattern Recognition 2023). Fine-tune VLMs (CLIP, BLIP, LLaVA, Gemma, Phi) under tight data constraints using LoRA, quantization, and knowledge distillation.
🔗 RAG & LLM Pipelines for Structured Extraction — Production RAG systems combining hybrid retrieval (dense + sparse), few-shot prompting, knowledge graph grounding, and calibrated confidence. Deployed with Ollama, Llama, Gemma, Phi, sentence-transformers, and ChromaDB — optimized for latency, memory, and privacy.
👁 Computer Vision & OCR for Documents — End-to-end OCR pipelines using RapidOCR, Tesseract, DocTR, and custom detection/recognition models. Scene text detection in low-resource scripts (Khmer), identity document fraud detection, layout analysis, and table extraction.
⚙️ Engineering & Infrastructure — FastAPI, React/TypeScript, Docker, WebSocket streaming, SLURM cluster management (40+ GPUs), MLOps with MLflow, CI/CD for ML pipelines.
📚 Research Output — 24+ peer-reviewed publications, 245+ citations, H-index 7. Ph.D. supervision of 5 students. PI/co-PI on 4 funded international projects (~€450K total). Program committee for ICDAR, ICPR, ICIP, Pattern Recognition.
Portfolio highlights:
1️⃣ DocAI — Production document extraction platform. 4 extraction modes, RAG retrieval, human-in-the-loop review, confidence scoring, anomaly detection. Exports to UBL 2.1 XML, EDI 810, CSV, JSON. 100% offline, zero cloud dependencies.
2️⃣ GlobalDoc: Cross-Modal VL Framework for Document Retrieval & Classification — WACV 2025 (IEEE/CVF). Cross-modal vision-language framework for real-world document image retrieval and classification.
3️⃣ VLCDoC: Vision-Language Contrastive Pre-Training for Cross-Modal Document Classification — Pattern Recognition 2023 (Elsevier, IF ~8). Contrastive VL pretraining for multimodal document understanding.
4️⃣ EAML: Ensemble Self-Attention Mutual Learning for Document Image Classification — IJDAR 2021 (Springer). Multimodal fusion using self-attention and mutual learning.
5️⃣ Multimodal Adaptive Inference with Anytime Early Exiting — ICDAR 2024 (Springer). Dynamic early-exiting strategy for efficient multimodal document classification — skip unnecessary layers when confidence is high.
6️⃣ Confidence-Based Knowledge Distillation for Low-Resource NMT — Applied Sciences 2025 (MDPI). Reducing training costs and carbon footprint via distillation + quantization.
7️⃣ DocSum: Domain-Adaptive Pre-Training for Document Abstractive Summarization — WACVW 2025. Domain-adaptive pretraining for generating document summaries.
8️⃣ IDTrust: Deep Identity Document Quality Detection with Bandpass Filtering — WACVW 2025. Quality detection and fraud prevention for identity documents.
9️⃣ LLMChain: Blockchain-based Reputation System for Evaluating LLMs — COMPSAC 2024 (IEEE). Decentralized evaluation and sharing.
Steps for completing your project
After purchasing the project, send requirements so Bakkali can start the project.
Delivery time starts when Bakkali receives requirements from you.
Bakkali works on your project following the steps below.
Revisions may occur after the delivery date.
Client sends 2–3 sample documents and target field list
I configure VLM extraction pipeline and run on samples