Beaver • 2025
Beaver | 4–5 day Procurement to ~10 mins
Deployed a custom RAG agent that enriched Beaver's 10,000-entry client database in 90 minutes versus 6+ months manually, cutting research overhead costs by 99%.
Human Impact
Zero headcount replaced; redirected 40 hours of manual data entry per week into high-level client strategy.
The Context
Beaver Process Equipment required a deep-dive enrichment of their 10,000-entry client database to identify high-value tender opportunities. In the industrial B2B sector, "who knows what" and "where projects are active" is the primary driver of sales, but this data is often buried deep within unstructured company websites.
The Problem: The Prospecting Bottleneck
The manual research process was a massive drain on resources. To manually review company profiles, map market positioning, and assess tender relevance for 10,000 records would take a dedicated team months of full-time work, costing an estimated $75,000+ in manual labour. By the time the data was ready, the tender windows had often already closed.
The Intervention: Razor Agent AI Pipeline
We deployed a custom RAG (Retrieval-Augmented Generation) Agent designed specifically for the industrial sector.
- Deep Crawl - The agent autonomously navigated target websites to find hidden project histories and service capabilities.
- Structured Extraction - Using Gemini 2.0 Flash, the system converted messy web text into clean, structured CRM data.
- Market Mapping - The tool automatically tagged each lead by commodity sector, geography, and tender relevance.
The Value Delivered
Instead of a $75,000 manual research bill and a 6-month wait, the entire 10,000-entry database was enriched and ready for outreach in under 90 minutes. The platform now functions as a "Forever Fresh" intelligence tool. Beaver can perform ad-hoc research on new prospects instantly, ensuring they are always first to the table for new tender opportunities at a fraction of the cost of traditional business development.
"Organisational transformation doesn't start with tooling; it begins by developing an internal culture."
— Nick Strine, Founder, SemiMassive
Responsible AI Practices
A core principle of this build was ensuring the client could trust and understand the AI's output. Every data point scraped by the agent was scored for accuracy based on how the information was sourced. Three tiers of data provenance were defined:
- Structured source - data found directly through a company's structured web pages, such as an about page or services listing. Highest confidence.
- General source - data gathered from unstructured content elsewhere on the company's website. Moderate confidence.
- External source - data retrieved via external search calls rather than the company's own site. Lowest confidence.
By surfacing these scores alongside every enriched record, the client had full clarity over how the AI reached its conclusions and could make informed decisions about which data to act on. This transparency turned the system from a black box into a tool the team genuinely trusted.
Technical Architecture
- AI Layer - Google Gemini 2.0 Flash (optimised for high-speed structured extraction)
- Vector Store - Pinecone Serverless (RAG-grounded search for technical specs)
- Intelligence - Google Search API + Trafilatura for deep-web crawling
- Output - Unified Sales Assistant mode with direct Excel/CSV CRM exports