Landing Page
Dashboard
Enquiry Form
Enquiry Page
Voice Note
Templates

QuoteReady: AI-Powered Trade Quote Readiness

Track 1: Improving an Existing Business Capability

Inspiration

Small trade businesses are often expected to quote from a short message, a photo, or a rushed call. Before committing to a fixed price, a tradie still needs to understand the actual scope, missing information, access requirements, and possible safety concerns.

Getting this wrong can mean underquoted work, wasted site visits, lost margin, and scope disputes. We did not want to build another generic chatbot, AI receptionist, or automatic quote generator. We built the step before the quote: a way to decide whether there is enough evidence to responsibly commit to a price.

For the hackathon, we focused on three plumbing templates—leaking taps, toilets, and hot-water issues—while keeping the system template-driven for future trades.

What it does

QuoteReady turns fragmented enquiries into a structured, reviewable job scope.

Accepts customer text, photos, call details, and spoken site notes
Extracts known facts, missing details, evidence, assumptions, and risk signals
Calculates an explainable quote-readiness score
Recommends: request more information, book an inspection, or prepare an owner-reviewed estimate
Creates editable customer follow-up drafts
Records approvals, evidence, scope changes, and audit events
Turns ElevenLabs voice notes into structured field evidence that can update the recommendation

AI interprets. Deterministic rules decide. Humans approve.

QuoteReady does not diagnose faults, set prices, create binding quotes, or automatically contact customers. Every price is operator-entered and every customer-facing action requires approval.

How we built it

We built QuoteReady with Next.js, TypeScript, Tailwind, Supabase, Gemini, LangGraph, ElevenLabs, Zod, Vitest, and Vercel.

Gemini 3.6 Flash processes enquiry text and photos in one multimodal request, extracting structured facts, evidence sources, certainty labels, and possible risk signals. Its output is treated as untrusted input and validated with Zod.

A deterministic rules engine—not the LLM—calculates readiness:

[ \text{Readiness} = 0.25D + 0.25E + 0.20A + 0.15C + 0.15R ]

where (D) is details, (E) evidence, (A) access, (C) confirmation, and (R) risk. Missing critical information, inspection conditions, and safety patterns can cap readiness or reroute the next action.

LangGraph orchestrates a fixed workflow:

Validate input → Gemini extraction → Zod validation → rules and safety checks
→ cited guidance retrieval → operator wording → scope version + audit event

We also created a curated 46-note service playbook. Notes are embedded with gemini-embedding-001, stored in Supabase pgvector, and retrieved through vector search with a keyword fallback. Retrieval is advisory only: it can improve guidance, but cannot change readiness, safety routing, or the recommended action.

ElevenLabs Scribe powers spoken site notes. A tradie can record an observation, review the transcript and fact changes, then apply it as a new immutable scope version. The updated evidence re-runs readiness, safety rules, and guidance.

Challenges we ran into

The biggest challenge was making AI helpful without allowing it to make a commercial or safety decision.

We solved this by separating AI interpretation from deterministic rules. Gemini can read messy enquiries, photos, and voice transcripts, but it cannot set a price, clear a safety issue, send a customer message, or decide that a job is safe to quote.

We also wanted ElevenLabs to be more than a microphone button. Spoken site notes now become structured evidence that can change a scope recommendation after human approval.

Accomplishments that we're proud of

Built a live end-to-end workflow from enquiry to approved follow-up and versioned field update
Created a deterministic, explainable readiness engine instead of an LLM-generated score
Added raw-text safety pattern routing, so important customer language is not lost in AI paraphrasing
Built cited retrieval that is technically real but isolated from high-impact decisions
Made ElevenLabs voice notes a core operational workflow, not a cosmetic feature
Added 10 labelled evaluation scenarios, 91 automated tests, and 96.2% statement coverage on the decision layer
Added per-run latency, token, and estimated-cost telemetry

What we learned

The best AI workflow is not always the most autonomous one. For high-impact business decisions, AI is strongest when it interprets unstructured information while deterministic rules and humans control repeatable, accountable decisions.

We also learned that a small, curated, cited knowledge base is more useful than a huge unverified corpus, and that voice becomes valuable when it captures field information that would otherwise never reach the system.

What's next for QuoteReady

Add route-level authentication, stronger tenant controls, rate limiting, and spend caps
Evaluate extraction using independently labelled real-world trade enquiries
Add web form, SMS, email, and field-service-platform integrations
Add reviewed playbook-management workflows
Expand verified templates to electricians, HVAC technicians, painters, landscapers, and other trades
Integrate approved scope packs with existing quoting, job-management, and invoicing tools

QuoteReady does not replace a tradie’s judgement. It makes the invisible parts of quoting—missing facts, assumptions, evidence, risk, and next actions—visible before a business commits to a price.