From demo-day crunch to a production AI estimating pipeline
How an Australian construction-estimating AI startup used one embedded Lyra engineer to get demo-ready, then rebuilt its plan-reading pipeline: faster processing, lower AI cost and CI cut from five minutes to 80 seconds.
An Australian AI startup that turns building plans into construction cost estimates had weeks until its accelerator demo day and a product that wasn't ready to show. One Lyra engineer started three days after being introduced, made the product demo-ready, then rebuilt the plan-reading pipeline. The founders have since committed to working with Lyra long term and asked for a second engineer.
At a glance
- Client: early-stage, accelerator-backed construction-estimating AI startup
- Need: demo-ready product fast, then a reliable estimating pipeline
- Team: 1 embedded engineer, a second requested
- Stack: TypeScript, Vercel AI SDK, Railway, trigger.dev, Stripe, PDF rendering
The problem
Demo day was close. The founders wanted tool calling in the AI chat, a design that no longer looked generic, and cleaner data in the database, all at once. Behind the demo, processing a full set of plans was slow and fragile.
What Lyra's engineer built
A chat that does real work. Moved the chat to the Vercel AI SDK with streaming and web search, added in-chat price research with citations checked on the server, and let the AI fill in customers' own Excel estimating workbooks.
A durable pipeline. Moved estimating onto background jobs with Railway and trigger.dev, so an estimate survives a page refresh or a restart, and logged the cost of every AI call.
A plan viewer. A PDF viewer with AI-detected items drawn over the plans, where shapes the user draws flow straight into the estimate.
Smarter plan reading. Close-up region extraction, measured beams and footings, and reading rules tuned per trade, all checked against a held-out set of plans the model has never seen.
Growth basics. A 7-day free trial on Stripe and an exit survey on every cancellation.
Results
- Engineer started three days after the introduction.
- Large commercial plan sets processed in under 7 minutes with 26% less AI cost.
- Document-type reading right on 39 of 40 unseen files, up from 33.
- CI from about 5 minutes to about 80 seconds, and about 10,000 lines of dead code removed.
- Free trial and cancellation survey live for the launch.
MVP development agency or an embedded engineer?
An MVP development agency gets you to a first version and leaves. The hard part for this startup came after the demo: making estimates accurate and fast on real plans. An embedded engineer who stays with the product, reviews the co-founder's pull requests and measures every change against unseen plans is how an MVP becomes something customers pay for.
Read more: What is a forward deployed engineer?
Frequently asked questions
