A self-hosted, enterprise-grade multilingual document intelligence and schema extraction platform that runs fully offline on private infrastructure.
RoleFull Stack Developer
DurationSept 2024 - Dec 2025
Client / PartnerPromeTechAi Innovations Pvt Ltd
IndustrySaaS / Enterprise Document Intelligence
Enterprises deal with massive volumes of unstructured documents, scanned PDFs, photographed reports, multi-column invoices, dense medical records, and multilingual government filings. Extracting usable data from these documents was entirely manual: operators visually scanned each page, manually entered data into spreadsheets, and cross-referenced tables by hand. This process was slow, error-prone, and unscalable. Many organizations also operated under strict NDA and data-residency rules, making third-party cloud AI services unsuitable.
Build a fully self-hosted AI-powered document intelligence platform that:
- Detects document layout structure
- Extracts multilingual text accurately
- Converts complex tables into structured HTML/JSON
- Supports schema-driven extraction
- Runs fully offline on private infrastructure
- Provides modern upload, viewer, chat, and export interfaces
Developed an AI-based document processing platform utilizing custom offline pipelines. The platform ingests documents natively, detects structural layouts (headers, tables, charts) using DocLayout YOLO, extracts text via multilingual PaddleOCR, and provides an intelligent semantic search interface through a Next.js frontend and Python FastAPI backend.
Built a two-product platform using Next.js 16 frontend and Python FastAPI backend:
- Product 1: Document Processor: Users upload PDFs or images. The backend runs a three-stage AI pipeline (DocLayout-YOLO for layout detection, PaddleOCR for multilingual OCR, and PP-StructureV3 for table extraction). The frontend provides a split-panel viewer with bounding boxes, overlays, extracted data inspection, and AI chat.
- Product 2: Schema Extraction: Users define schemas or auto-generate them using AI. The system extracts structured values matching schemas, allows editing, exports JSON/CSV/XML, and routes data into APIs, databases, email, or cloud storage.
- Fully offline AI processing pipeline with no dependence on third-party APIs.
- Automated async document upload and batch processing queues.
- Bounding-box structural layout grounding (paragraphs, tables, images) using DocLayout YOLO.
- High-accuracy multilingual OCR pipeline powered by PaddleOCR.
- Reading order and spatial layout preservation for complex documents.
- Bordered and borderless table extraction transforming images to structured JSON format.
- Component-level image and chart extraction for downstream analytics capabilities.
- Standardized AI SDK for seamless frontend and backend integration.
- Intelligent NLP-powered semantic search, document classification, and entity recognition.
- Interactive React-based Next.js frontend for analyzing and visualizing layout-grounded results.
- Designed and built the interactive Next.js components for document uploading, table visualization, and layout result rendering.
- Orchestrated the Python FastAPI microservices to handle large asynchronous document queues without blocking the main event loops.
- Implemented and tuned DocLayout YOLO models to physically map document structures into strict bounding-box JSON coordinates.
- Integrated PaddleOCR pipelines within the bounded layout boxes to perform focused high-accuracy text extraction.
- Developed custom extraction algorithms converting complex spatial table grids (both bordered and borderless) into coherent 2D JSON data structures.
- Created a standardized AI SDK serving as the programmatic bridge unifying the ML pipelines, React frontend, and database ingestion workflows.
- Built the global smart search UI, allowing instant semantic retrieval of information across thousands of scanned enterprise documents.
- Heavy memory usage for PDFs during layout analysis.
- Long-running inference requests causing client connection timeouts.
- Duplicate detections from object detection models overlapping on text blocks.
- Reading order reconstruction in complex multi-column layouts.
- Borderless table extraction complexity.
- Responsive overlay rendering accuracy on varying screens.
- Large ML Docker image size making deployments slow.
- Proactive
psutil memory monitoring with thresholds to control queue flow.
StreamingResponse with keep-alive heartbeat tokens to keep connections open.- IoU-based duplicate filtering to clean up overlapping bounding boxes.
- Coordinate sorting for reconstruction of accurate reading order.
- Dedicated table pipeline focused solely on extracting tabular cells.
- Percentage-based responsive overlay positioning.
- Multi-stage Docker builds reducing image size.
- Reduced processing time from 15–20 minutes of manual work to seconds/minutes automated.
- Supported PDFs up to 50MB / 20 pages.
- Prevented OOM crashes using proactive memory controls.
- Improved OCR accuracy via upscale preprocessing.
- Enabled 300-second processing windows without timeout.
- 100% private on-premise inference.
- Built proprietary document AI IP.
- Opened enterprise markets requiring strict privacy compliance.
- Enabled SaaS pricing tiers and monetization models.
- Reduced manual operational costs for client industries.
- Enabled non-technical users to create extraction templates easily.
Frontend (Next.js) handles UI, uploads, workflows, and AI chat. Backend (FastAPI) runs OCR and extraction pipelines. MongoDB stores history and schemas. Supabase manages auth and credits. AWS S3 + CloudFront store and serve assets. Dockerized deployment across environments.
- Streaming responses instead of full buffering.
- Page-by-page PDF processing.
- Threaded CPU-bound inference.
- Parallel uploads to S3.
- Configurable limits via environment variables.
- Smaller Docker runtime images.
- Frontend API proxying.
- Extensible modular workflow engine.