Searchable Archive IA Engine for a European Legal Tech Provider

Devox’s partnership with the U.S. industrial group began after a board-level referral: earlier document-automation results convinced the client to invite us in to unlock its decade-old scan archive. The mandate was clear — deliver full-text search and audit-grade traceability without altering the company’s heavily customised on-prem ERP.

We designed a sidecar Document Intelligence Layer: a Dockerised FastAPI service that tails the ERP’s export queue, cleans images, auto-detects English or Spanish, and feeds a self-trained Tesseract OCR. A 48-hour pilot on 5,000 pages hit 94% accuracy and sub-second Elasticsearch search, securing go-live approval.

Now every new scan runs a single pass — enhancement, language detection, confidence-scored OCR, live QA on low-certainty tokens, PDF/A sealing, and instant indexing — and writes straight back through the ERP’s HeadersRepository/RowsRepository. The once-static archive is fully searchable, audit-ready, and remains entirely on-prem.

• Solution Architect (document intelligence & legal DMS integration)
• 2 Backend Engineers — Python/FastAPI, .NET SDK wrapper
• QA Automation Engineer
• DevOps Engineer (on-prem Docker Swarm, CI/CD, observability)
• Business Analyst (legal workflows & data-privacy compliance)
• Project Manager

• Image-only archive. Roughly 2 million legacy contracts were stored in TIFF/PDF format without a text layer, forcing the search engine to index only filenames and metadata, while the actual clauses remained invisible.
• Manual clause retrieval. Paralegals spent 800+ hours every month retyping passages for discovery requests — one page at a time — creating a 7-day backlog for even routine keyword searches.
• Multi-language corpus. Documents arrived in English (55%), German (30%), and French (15%), but language indicators were missing, so every OCR pass had to auto-detect the script before applying the correct dictionary and hyphenation rules.
• Variable scan quality. Source material ranged from 300 dpi PDF/A to 150 dpi fax scans with skew and shadowing, which dropped baseline OCR accuracy below 85% and triggered costly manual clean-ups.
• Strict on-prem residency. Client policy and GDPR constraints prohibited any external API calls; the entire AI stack, including model updates, had to run behind the firewall with verifiable data isolation.
• Sub-second search SLA. Legal teams demanded a < 1s response across the whole corpus, meaning the pipeline had to deliver both OCR throughput (10,000 pages/day) and near-real-time indexing without adding hardware.
• Confidence blind spots. Early pilots returned plain text without certainty scores or source-page anchors, leaving reviewers unsure which values were accurate and undermining audit defensibility in court.

• Frontend: React 18 with Vite, i18next for multilingual UI, React Router for routing logic, and Ant Design for internal QA console components.
• Backend & Orchestration: FastAPI (Python 3.11), .NET 6 wrapper for enterprise-grade OCR SDK, Pydantic v2 for schema validation, and Celery for background job orchestration.
• Pre-processing & AI: NeoML filters for deskewing and denoising, spaCy for named-entity recognition and clause tagging, with Tesseract fallback for low-resolution or degraded scans.
• Search & Storage: Elasticsearch 8.x for real-time full-text indexing, PostgreSQL 14 for metadata and audit logs, NFS-attached storage for both originals and standards-compliant PDF/A outputs.
• DevOps & Deployment: Dockerized microservices deployed to an on-prem Docker Swarm cluster (3 manager nodes + 6 workers), GitLab CI/CD for pipeline automation, and Ansible for infrastructure as code (IaC).
• Authentication & Security: SAML SSO integrated with Microsoft Active Directory, JWT service tokens for API access, on-premises HashiCorp Vault for secrets management, and TLS 1.3 encryption across all internal services.
• Monitoring & Logging: Prometheus exporters and Grafana dashboards for system health, Loki for centralized log collection, and Alertmanager for real-time incident notifications.

Devox Software delivered a fully on-prem Searchable Archive IA Engine that integrates high-precision OCR into the client’s document management system, combining microservice orchestration with intelligent automation.

• Single-pass ingest pipeline. Every new TIFF or image-PDF dropped into the watch directory triggers a FastAPI-based orchestrator that enhances image quality (deskew, denoise), detects the dominant language, and routes the page to an OCR engine via a .NET interface.
• Language-adaptive OCR engine. For each page, the service dynamically selects an English, German, or French processing profile, applies legal-domain dictionaries, and assigns a confidence score to each extracted token.
• Confidence-aware validation layer. Low-certainty fields (under 95 %) are flagged for review in a React QA console. Corrections are logged once and instantly reinforce lightweight language profiles.
• Instant PDF/A packaging. Cleaned documents are merged with an invisible text layer, wrapped in a digitally signed PDF/A-3b container, and archived with SHA-256 hashes for evidentiary integrity.
• Real-time search indexing. Extracted text and enriched metadata are piped directly into Elasticsearch, enabling sub-second full-text search across the entire archive and meeting discovery SLAs for legal and compliance teams.
• Full traceability. Each search result traces back to its source document, OCR version, confidence metrics, and reviewer corrections, delivering audit-ready lineage across every clause.
• Secure on-prem deployment. The engine runs in an air-gapped environment within a six-node Docker Swarm cluster. GitLab CI/CD deploys verified containers; Prometheus and Grafana provide live observability, while all data stays within the firm’s firewall, satisfying strict residency policies.

BUSINESS OUTCOMES

• 72% reduction in manual effort. Paralegals now review nine documents per hour, instead of just two, freeing up two full-time roles for higher-value legal work.
• Sub-second discovery. Keyword searches now resolve in ~0.4 s across the full archive, even at peak load.
• 98% archive coverage in 8 weeks. Nearly all legacy scans were converted into PDF/A documents with full-text indexing and evidentiary traceability.
• 99.2% text-level accuracy. Confidence scoring and human-in-the-loop corrections reduced OCR misreads by 63 %, meeting defensibility thresholds for court and audit review.
• One-year ROI. Reduced retyping hours and faster case prep delivered full payback in 11 months.

TECHNICAL OUTCOMES

• 10,000-page daily throughput on existing hardware. The pipeline peaks at 14 pages per second with a CPU usage below 65%, meeting the client’s growth forecasts.
• Modular micro-architecture. React QA console, FastAPI orchestrator, OCR workers, and indexing service deploy independently, enabling same-day hot-fixes without downtime.
• Live confidence tracking. Every token stores a score and a source-page anchor; low-certainty zones are surfaced in the UI and feed into nightly model fine-tuning.
• On-prem, zero-egress security. All containers run inside a six-node Docker Swarm; no external APIs, complete data isolation, and audit-ready PDF/A-3b output.
• End-to-end observability. Prometheus & Grafana dashboards monitor ingest latency, OCR accuracy, and index freshness; SLA breaches trigger instant Slack alerts.