URL IntelligenceCost-Aware ExtractionBrowser Automation

Vergabepilot.AI

Autonomous tender data extraction across changing procurement portals

Vergabepilot.AI automates tender discovery across changing procurement portals. Its cost-aware pipeline moves from reusable scrapers to browser automation and AI extraction, remembers successful routes, and retrieved 97% of 100 German tenders for only $0.02 in model costs.

Vergabepilot.AI workflow from tender discovery through cost-aware AI processing to decision-ready results

In collaboration with

Tender extraction walkthrough · portal detection, adaptive retrieval, and document processing

Benchmark table comparing model extraction success, latency, and cost per run — FigureExtraction success, latency, and cost across nine language models.

Model evaluation

Multi-model benchmarking

Nine models were compared using extraction success, response latency, and cost per run. Gemini 2.5 Flash Lite achieved the strongest overall balance in this benchmark.

Best model: Gemini 2.5 Flash Lite
Extraction success: 83%
Latency: 4.176 ms
Cost per run: 0,00181 $

How the extraction pipeline works

Each tender URL first passes through a URL intelligence layer, which identifies the portal type and selects the most suitable extraction order. The system then tries different strategies — from cached and deterministic methods to adaptive browser automation, LLM-generated scrapers, learned routes, and agent-based navigation.

The strategies are executed from cheaper and faster methods to more advanced ones, and the process stops as soon as one succeeds. The retrieved tender documents are then downloaded and processed to extract structured procurement information for further use or reporting.

Cost-aware tender extraction pipeline from URL intelligence through reusable extraction strategies to structured procurement data — FigureCost-aware extraction from tender URL to structured data.

Technical focus

URL IntelligenceCost-aware extractionPlaywright automationLLM-generated scrapersSelf-healingComputer Use AgentDocument extractionScraper caching

Implementation details

Portal-specific strategy selection reduces unnecessary processing and AI calls
Failed LLM-generated scrapers can automatically be corrected and retried
Successful scrapers and navigation routes are stored and reused for future tender URLs

← Back to All Projects