From 4-hour bids to 5-minute automation

The Problem

A construction company was drowning in their bid process. Every day, dozens of RFI (Request for Information) emails arrived with attached bid documents — plans, specifications, material lists. Each bid required manual review, classification, and data extraction before the estimating team could even begin pricing.

The process consumed roughly 4 hours per bid. At ~2 bids per day, that’s a full-time person doing nothing but document triage. The bottleneck wasn’t estimating — it was the administrative overhead of organizing incoming work.

Critical information was scattered across email threads, Google Drive folders, and spreadsheets. Bids were missed. Deadlines were blown. The team knew there had to be a better way, but every “solution” they’d tried just moved the manual work from one place to another.

The Solution

We built two interconnected systems that transformed the entire bid pipeline:

bid_manager handles the email side — monitoring incoming RFIs, classifying emails using Claude 3.5 Sonnet, extracting key metadata (project name, deadline, scope, materials), and organizing everything into a structured pipeline in Google Sheets and Drive.

solid-surface-assistant handles document analysis. It builds a knowledge graph from bid documents, enabling the team to query across projects: “Show me all countertop specs from the last 6 months” or “What’s the typical timeline for a hospital renovation?” This isn’t just search — it’s institutional memory.

The two systems work together. When bid_manager classifies an incoming email and extracts documents, solid-surface-assistant automatically indexes them into the knowledge graph.

Results

The numbers speak for themselves:

• 90%+ time savings on bid processing — from 4 hours to under 5 minutes
• >95% classification accuracy on email categorization
• 406 tests ensuring system reliability
• Phases 1-2 delivered — email automation and document analysis are in production
• Phases 3-4 on roadmap — automated takeoff estimation and proposal generation

The system processes bids faster than the team can estimate them, which is exactly the right bottleneck to have.

How We Built It

All six methodology steps applied to this project:

1. HCI / Wireframing — We spent two weeks observing the actual bid process. The documented workflow had 6 steps. The real workflow had 23. We mapped every workaround, every Excel macro, every “I just remember to check that.”

2. SpecKit / Specify — Every email classification category, every extraction field, every edge case (What about emails with no attachments? What about forwarded chains? What about attachments in wrong formats?) was specified before building.

3. Pure Functions — The classification pipeline is a chain of pure functions. Email in, structured data out. No side effects. Each stage is independently testable and replaceable.

4. TDD — 406 tests. Classification accuracy tests. Extraction validation tests. Integration tests. Edge case tests. The test suite runs in under 3 minutes and catches issues before they reach production.

5. Regression Testing — When we updated from Claude 3 to Claude 3.5 Sonnet, the regression suite caught 3 classification edge cases that would have silently degraded accuracy.

6. AI CI/CD — Automated deployment with evaluation pipelines. When classification accuracy drops below threshold, the system alerts before users notice.