Overview

The International Labour Organization (ILO), a specialized agency of the United Nations, needed a faster, more consistent way to process and act on a high volume of publicly available documents. The manual workflow was time-consuming, difficult to scale, and dependent on individual reviewers reading and assessing each document by hand.

To solve this, SolDevelo designed and built an end-to-end document intelligence platform that automatically discovers, ingests, analyzes, and classifies these documents using a combination of modern web automation and large language models (LLMs). The system extracts information and it makes structured, explainable relevance decisions that help ILO staff focus their attention where it matters most, and it puts an AI assistant at their fingertips to explore the underlying documents conversationally.

The platform is currently running as a pilot within ILO, where early feedback has been strongly positive.

The Challenge: Manual Workflows vs. Exploding Data Volume 

ILO’s teams were spending significant effort manually locating relevant documents, reading through them, and judging whether each was worth acting on. This created several problems:

• Volume and velocity: Documents were being published faster than human teams could consistently review them. 
• Subjectivity: Relevance judgments naturally varied from reviewer to reviewer, making the decision-making process difficult to audit or standardize. 
• Slow time-to-insight: By the time a document was assessed, the window of opportunity to act on its insights could already be closing. 

The goal was a system that could take on the heavy lifting of discovery, extraction, and first-pass decision-making while keeping a human firmly in control. Crucially, the platform had to earn the trust and adoption of the people using it daily. 

Our Approach: Balancing Technical Scale with User Adoption 

We treated this as two intertwined problems: a technical one (build an accurate, scalable AI pipeline) and a human one (make sure people actually adopt it). From the start, ILO told us that user adoption was as important as raw capability – a brilliant system nobody trusts or uses delivers no value.

So we built the platform as a set of focused microservices communicating asynchronously through a message broker, allowing each stage – scraping, orchestration, classification, storage – to scale and evolve independently. In parallel, we worked directly with ILO to design the product around how their people actually work.

Our approach delivers a dual advantage across the organization:

• Technical Resilience: Individual components can be upgraded, scaled, or replaced independently without risking system-wide downtime.
• Operational Stability: It guarantees a cost-effective infrastructure capable of absorbing massive data spikes smoothly, keeping the platform fast and reliable for end-users.

Technology Stack

The platform is composed of independently deployable services orchestrated with Docker:

• Frontend: React 19 + TypeScript, built with Vite, TanStack Router & Query, shadcn-style UI components (Radix UI primitives + Tailwind CSS), with OIDC-based authentication.
• Management backend: Spring Boot (Java 21) with PostgreSQL, handling orchestration, persistence, and the public API.
• Classifier service: Python with a Haystack-based AI pipeline driving the LLM extraction and decision logic.
• Scraper service: Spring Boot with Playwright browser automation and robust HTML/document parsing.
• Chat server: a TypeScript/Express service powering the in-app AI assistant.
• Messaging & infrastructure: RabbitMQ for inter-service communication, Redis for caching, Keycloak for identity and access management (OIDC/OAuth2), and a CI/CD pipeline deploying containerized images to AWS.

This service-oriented architecture means each part of the system can be scaled, replaced, or upgraded without disrupting the others – including swapping out the AI models behind the scenes.

Scraper Technology: Transforming Messy Web Content into Clean AI Inputs 

Reliable data is the foundation of any AI system. Traditional web scrapers often break when trying to pull data from complex, interactive web pages.

To prevent this, our scraper service uses Playwright for full browser automation. The system easily navigates JavaScript-heavy pages that simpler scrapers cannot read. Messy real-world HTML and downloaded documents are automatically parsed, cleaned, and transformed into structured, standardized text ready for immediate AI analysis.

Because scraping runs as its own service and publishes results onto the message bus, ingestion can scale and run on schedules independently of the rest of the system.

AI Strategy: Optimizing Cost, Flexibility, and Trust 

The intelligence of the platform is its defining feature, and we engineered it deliberately rather than reaching for a single all-purpose model.

1. Multiple LLMs, Matched to the Job

Not every task needs the most powerful (and most expensive) model. We architected a multi-model approach:

• Lighter, faster models handle structured data extraction – pulling key fields and facts out of each document.
• Stronger reasoning models handle the decision-making – judging how relevant a document is and whether it matters to ILO specifically.

This division of labor gives ILO the best of both worlds: high accuracy on the judgments that matter, with efficient, cost-effective processing for the routine work.

2. Eliminating Vendor Lock-In with OpenRouter 

Rather than wiring the system to a single AI provider, we route all model calls through OpenRouter as an LLM gateway. This is a strategic advantage for ILO: they can freely switch between models and providers as the AI landscape evolves, optimize for cost or quality on a per-task basis, and avoid being locked into any single vendor – all without code changes.

3. Trustworthy Answers via a RAG Pipeline 

Beyond classification, we built a Retrieval-Augmented Generation (RAG) pipeline and an integrated chatbot directly into the application. Users can ask questions in natural language and get grounded answers, because the assistant’s context is built from a specific, curated set of documents rather than the open internet. This keeps responses relevant, accurate, and anchored to ILO’s actual source material, turning a static archive into something users can interrogate conversationally.

A Modern, Accessible Frontend

The interface was built to feel modern, fast, and effortless. Using a shadcn-style component system (Radix UI + Tailwind CSS), we delivered a clean, consistent design that is fully responsive across devices and built with accessibility as a first-class concern, leveraging the accessible primitives and keyboard/screen-reader support that the component foundation provides.

The result is a tool that feels professional and polished, lowering the barrier for new users and reinforcing trust in the system.

Designing for Adoption: Working Hand-in-Hand with ILO

Because ILO emphasized user adoption from day one, we made it a core part of the delivery.

We helped ILO organize structured user testing inside the organization, where a selected group of staff were invited to use the tool and share their impressions. To make that feedback loop as smooth as possible, we built dedicated feedback views directly into the application. Instead of asking users to switch to separate documents, spreadsheets, or external forms, they could submit feedback in-context, without ever leaving the tool.

This had a double benefit: it removed friction (so we got far more, and far more honest, feedback), and it made users feel ownership of the product they were helping shape – a powerful driver of adoption.

Challenge Solved: Finding the Optimal Combination of LLMs 

One of the bigger challenges was model selection. The “right” LLM is not obvious in advance, and the wrong choice means either poor accuracy or unnecessary cost. We ran extensive evaluation across many candidate models, measuring their performance on both extraction and decision tasks. That hands-on experimentation is what led us to the multi-model architecture, pairing efficient models for extraction with stronger models for relevance and decision-making, and the OpenRouter gateway is what lets ILO keep refining those choices over time.

Business Outcomes & Current Status

The platform is now running as a pilot within ILO, and the early feedback has been very positive. While the pilot is intentionally scoped to a subset of the organization rather than ILO as a whole, the impact is already visible: it is changing how participating teams approach their day-to-day work and, importantly, how they make decisions about the documents they handle.

• Drastic Efficiency Gains: Automated ingestion and first-pass assessments have significantly reduced the time spent on manual document sorting, freeing teams to focus on strategic action rather than reading piles of text.
• Instant Knowledge Retrieval: Staff can query a massive database of documents conversationally, extracting hidden insights in seconds rather than hours of manual searching.

By taking the heavy lifting out of data processing, the platform is empowering the ILO’s workforce to spend less time managing data and more time acting on it.

Looking for a Reliable Technical Partner?

From modern frontend design to complex backend orchestration and AI pipelines, we deliver enterprise-grade software engineered for high organizational adoption.