Construction Estimation Platform
QPOI designed and built the AI engine at the core of a production construction SaaS platform: a multi-stage pipeline that reads blueprints and produces complete, reviewable project estimates, plus the ten-step project lifecycle that carries every job from initiation to closeout.
The challenge
General contractors spend days reading dense blueprint sets by hand to price a job. The manual process is slow, inconsistent between estimators, and hard to scale: every bid starts from a blank page.
Off-the-shelf AI can't be trusted with money. A construction estimate has to be traceable, reviewable, and internally consistent across every document it feeds: quotes, contracts, schedules, budgets. A black-box answer is worthless the moment a client asks where a number came from.
- The full AI analysis engine: a 30-plus stage pipeline from blueprint PDFs to a complete project estimate
- The ten-step project lifecycle, from initiation through estimation, bidding, and construction to closeout
- The real-time, marketplace, team, and performance infrastructure around it, in production across Angular and .NET
The AI estimation engine
QPOI built the analysis pipeline end to end: every stage, the orchestration around it, and the review tooling that makes the output trustworthy enough to bid with.
A 30-plus stage pipeline reads blueprint PDFs and produces a full project estimate: bills of materials and labour breakdowns across 16 trades, cost rollups, an executive summary, procurement and daily construction schedules, permits and compliance, environmental lifecycle reporting, value engineering options, and a client-ready quotation package.
The original sequential chain was re-architected into a dependency-graph fan-out: independent trade analyses run in parallel, then dependent stages execute in ordered waves. A validated production run dropped from 23 minutes to about 12, with outputs verified against the sequential baseline.
Estimate figures live in a canonical takeoff store where recalculation is arithmetic, not AI. When a user edits a line item, the change cascades through every downstream document: reports, executive summary, quotes, bid packages, and the generated contract. Even spec changes propagate into narrative prose, so swapping 3,000 PSI concrete for 4,000 PSI updates the contract text too.
AI output is never blindly trusted. Reviewers get estimate version snapshots, per-item replacement with grounded AI suggestions, and staleness tracking that flags which reports an edit has invalidated, with a one-click parallel refresh that regenerates only those. Reference available upon request.
Wider platform engineering
Beyond the engine, QPOI built the platform around it: the systems that turn an analysis pipeline into a product construction teams use every day.
Designed and built the real-time layer: two SignalR hubs, one broadcasting analysis progress and workflow status, the other streaming AI responses token by token, so users watch long-running operations instead of waiting on them.
Profiled and rebuilt the hot paths on both sides of the wire: a heavy calendar session went from 654 backend queries to 254, and the estimation screen from 53 API calls to 27. Request coalescing and caching on the frontend, identity-resolution caching on the backend, and Web Workers keeping report compilation off the main thread.
Built the subcontractor marketplace: map-based search and filtering over live job data using SQL Server geography types and clustered Google Maps rendering, so trades find relevant work near them.
Designed and built team-based access: roles and seat management aligned to how construction companies actually operate, with tokenized invitation flows for workspace onboarding.
Per-trade bid packages generated from the estimate, trade award flows, and AI-drafted contractor-client contracts that stay consistent with the reviewed takeoff figures instead of quoting stale numbers.
Budget tracking synced from the takeoff, timeline and task scheduling with a project calendar, and an in-app AI assistant that answers questions over the project's own context.
Business impact
Estimates that took days of manual blueprint review are generated in about 12 minutes, roughly half the pipeline's original runtime after parallelisation
One edit updates every document: quotes, contracts, and reports stay consistent with the reviewed estimate, with no manual reconciliation
Hundreds fewer backend queries per session and measurably faster screens across the platform
Delivered end-to-end and in production: AI pipeline, frontend, backend, real-time, and team infrastructure
The pipeline stage by stage, how the parallelisation proved it kept the same numbers, and the faults the validation gates caught before users could.