Home / Cases / Construction PM Workbench
Commercial Construction · Sydney

An AI workbench for construction PMs: where the chaos of site meets the slowness of the office

A construction PM's day disappears into four things: site WeCom messages, contracts and variations, takeoffs and quotes, and reporting upward. We built a PM-centric AI workbench — site messages auto-structured, AI-flagged contract risk, change management with audit trail, Procore API integration.

IndustryCommercial / Custom-resi build Size20–80-person builder Timeline6-week MVP StatusPilot
60%
Contract review time
8h/wk
PM admin saved
100%
Audit trail coverage
5
Integrated modules

About the Client

A Sydney commercial builder running 5–8 jobs (residential + small commercial) with a 20–80 person team. They already use Procore as primary PM software, but site comms run on WeCom / WeChat, email is Outlook, contracts live in SharePoint. Procore's API sits unused — PMs spend their days as human glue between systems.

Their Challenge

Why ManifoldX

Our founder Banruo came up through Sydney sites himself — 3.5 years as Site Engineer + Foreman + 7 months as Junior PM. He knows which pains are real and which are coder fantasy. Two construction software vendors had pitched the client; neither understood why PMs must keep using WeCom and can't be forced into Procore. Their proposals didn't land.

The Solution

Don't replace Procore — sit on top of it as a 'PM-perspective workbench'. 5 modules:

1. Site message structuring engine (site-record-ai)

Pulls WeCom / WeChat group messages (text / image / voice) in real time. AI auto-classifies into 6 categories: defects / daily logs / RFI / variations / safety / delays. Each entry timestamped, attributed, project-linked. Two outputs: PostgreSQL (machine queryable) + Obsidian Markdown vault (PM readable + annotatable + drawing-linkable). Bot drops a confirmation card in the group before each entry is finalized.

2. Contract review workbench

Drop a subcontract / variation / purchase order PDF in. AI extracts 15 key fields (parties, amounts, payment milestones, scope, special clauses…), compares against template and historical contracts, flags 'abnormal clauses' (off-market amounts / missing standard clauses / ambiguous wording), and generates redline suggestions. PM finishes review and one-clicks the result into Procore's Commitments module.

3. Variations & risk ledger

Every variation (verbal owner ask / email / on-site call) is forced through a 4-step flow: evidence → quote → both-side sign-off → account entry. AI scans emails / WeCom / site recordings to flag 'this looks like a variation' and prompts the PM to start the flow. At final account, exports a full evidentiary package in one click.

4. Takeoff & pricing assistant

PDF drawings → AI summary + previous-version diff → semi-automatic takeoff (high-volume low-risk items auto, critical items flagged for human review) → pulls comparable historical jobs as price baseline. Cuts quoting from 2 weeks to 3–5 days.

5. Procore bi-directional sync

Read-heavy, write-light: pull Procore projects / commitments / RFIs / daily logs in for AI enrichment; only PM-confirmed structured items get written back (so the primary system stays clean).

Tech stack

FastAPI + Python PostgreSQL OpenAI GPT-4o Anthropic Claude (合同分析) WeCom Webhook Procore API Obsidian (本地 vault) RQ + Redis (队列)

Working with us

Weeks 1–2: site-record-ai core live, receiving authorised messages from 3 site WeCom groups, AI-classifying, writing to Obsidian. Weeks 3–4: contract review workbench v1, subcontracts only. Week 5: variations ledger. Week 6: Procore bi-directional sync done. A Senior PM acted as product owner, 2 hrs of review weekly.

I used to spend 60% of my day moving information from system A to system B. Now AI does that, and I get to actually be a project manager again — visit site, meet clients, make decisions. — Pilot PM (paraphrased)

Impact

What's next

Pilot feedback positive across 3 sites; rolling out to all 8 from June. Next phase: BIM / IFC model integration (paired with our Revit Takeoff tool) and 'voice-to-report' (PM dictates in the car, AI ships an evening report to the owner). This is also the direction we're targeting for SaaS, aimed at 20–200 person builders.