Proof of work

Systems that hold up
in production

I don't consult in theory. I've built and operated a live automation suite for a major Sydney remedial concrete & building repair contractor — a full production environment with real stakeholders, real data, and real consequences. Here's what that looks like.

1,226CRM opportunities migrated, cleaned & structured
12Automated production workflows built & deployed
767+Automated tests in the production suite
1,099Tenders backfilled and classified in a single day
Interactive Systems Simulation

Live Tender Pipeline Runner

Watch a real Sydney commercial tender move through ingestion, AI schema extraction, human safety sign-off, and production writes.

01 Email Intake
02 AI Extraction
03 Human Gate
04 Multi-System Sync
Inbound TriggerEmail Server · Sydney NSW
Narrow AI Parser & HITL GatePydantic v2 · OWASP LLM01

Click ▶ Run Automated Pipeline above to simulate automated extraction and verification.

Production Multi-SyncIdempotent API Writes
📁Dropbox Folder StructurePending
💼GoHighLevel CRM OpportunityPending
📅Google Calendar BookingPending
📊Tender Register SheetPending
🛡Audit Log & Rollback ScriptPending

The scope

What was built

A 12-module Python automation suite running on a production server in Sydney — connecting email, CRM (GoHighLevel), Dropbox, Google Calendar, and Google Sheets into a single end-to-end tender operations pipeline, with AI judgment at every key decision point.

Production Architecture

12-Module End-to-End Tender Automation

Live in Sydney Production
01Deterministic

Enquiry Intake

Email & Webhooks

Inbound tender emails parsed, sender extracted, duplicate checked against CRM.

02Strict Schema

AI Scope Extraction

Claude 3.5 + Pydantic

Scope of works, tender due date, and job metadata extracted with strict schema.

03Zero Unchecked Actions

Human Safety Gate

HITL Verification

Ambiguous dates, duplicate names, or edge cases trigger Telegram/email confirmation.

04Idempotent Writes

Production Sync

Dropbox · GHL · Calendar

Folders auto-created, CRM opportunity logged, and site inspection booked.

Step 01 Architecture Details:

Deterministic Python listener extracts attachments (PDF, DOCX) and deduplicates company names against existing 309 company entities in GoHighLevel.

Email Intake

Tender enquiry emails are automatically extracted, duplicate-checked against CRM and register, and converted into CRM opportunities with human confirmation when doubt exists.

Tender Folder Creation

Standardised Dropbox tender folders auto-created with correct naming convention and template structure — no manual setup per job.

Tender Review Summary

A one-page bid/no-bid assessment document auto-generated from the enquiry email, embedded with site photos and scope summary.

Calendar Booking

Emails naming a firm date are automatically classified as site inspections, reviews, or tender due dates, and booked on the correct calendar with colour-coding and reminders.

Follow-Up Cadence

Submission review reminders auto-created 2 working days before tender due dates, with scope-specific checklist items drafted by AI.

Estimate Draft

Scope of Works documents are read by AI and a rough task list is auto-patched into the client's real Excel estimate template (VBA, macros and formatting preserved).

Handover Automation

Sales-to-projects handover pack auto-assembled: handover sheet, budget workbook, and 8 standard subfolders — verified against real job data.

Job Won Workflow

On CRM status change to Awarded: management notified, engineer nominated via free-text email reply (interpreted by AI), handover pack flagged, and meeting booked.

Pipeline Health Report

Replaces the CRM's native funnel dashboard (which was found to silently mix two incompatible definitions of 'Won' — a 9x discrepancy). Three clean, unambiguous metrics.

Register Sync

One-way Tender Register to CRM sync for status and due dates — keeping the system of record current without manual double-entry.

Weekly Reporting

Weekly 'Work Won' statistics compiled from live CRM pipelines and drafted into an internal summary email.

Smoke Test

End-to-end live-fire test: borrows real scope content, runs the full pipeline, then rolls everything back cleanly. Used for deployment validation.

Engineering discipline

Built to hold up under pressure

Deterministic-first

Every repetitive action is a Python script. AI (Claude) is used only for genuine judgment calls — exactly one narrow, schema-validated LLM call per feature. No free-text, no guesswork.

Human approval gates

No AI output acts commercially without human review. Company dedup, email sends, priority classification — all gated behind explicit sign-off or confirmation.

767+ automated tests

Including adversarial input tests against prompt injection (OWASP LLM01). Untrusted email content is treated as data, never as instructions.

Production-hardened

DigitalOcean Sydney, per-run file locks to prevent race conditions, structured JSON logging with correlation IDs, rollback scripts for every write operation.

What this means for you

This is the standard you get

When you engage Finder Technologies on a fractional advisory basis, this is the level of engineering, discipline, and production-readiness you receive. Not theoretical consulting — actual systems that your team can rely on from day one.