QuoteIQ

Inspiration

Industrial buyers send requests for quotes as messy emails, PDFs, and spreadsheets, and the same steel sheet can be written five different ways. Experienced reps decode each line by hand, which is slow and error-prone, and a wrong grade or gauge can mean a rejected shipment or a quote that loses money. We wanted an AI system that speeds up quoting without making the one mistake this work cannot afford: guessing.

What it does

QuotePilot AI turns a messy RFQ into a ready-to-review quote. It extracts each line into a standard spec, matches it to a supplier catalog with a confidence score, prices it against benchmarks with a short explanation, and flags compliance needs such as mill test reports and certifications. Any line it cannot resolve, such as a missing grade or a gauge that contradicts the stated thickness, goes back to the buyer as a specific question and is never quoted. A rep approves the result and downloads a PDF with a per-line audit trail.

How we built it

Google ADK orchestrates the pipeline as a graph of agents.
Nicrron is the model gateway, so every model call goes through its OpenAI-compatible endpoint.
Block Convey Prism traces each run as one session per RFQ, covering model calls, tool calls, guardrail results, and the gate’s decision for every line.
Deterministic code handles unit and gauge conversion, catalog matching, pricing, and compliance rules. The model only reads and extracts text, so it cannot set a price.
Guardrails in code screen RFQ text for injected instructions, redact personal data, enforce a confidence threshold and a margin floor, and block conflicting specs.
Ninoo Builder for the front end: an upload screen, a line-by-line review table, clarification prompts for flagged lines, and a PDF download.
Python with [FastAPI] for the backend, exposing the pipeline as an endpoint the front end calls.
Claude for generating the PDF quote with its per-line audit trail. -Ninoo for hosting.
Gmail API with read-only OAuth for pulling RFQ emails.

Challenges we ran into

it took us some time to figure out the Ninoo builder setup

Accomplishments that we're proud of

we were able to build a working prototype pretty quick
We tested failure cases deliberately: an unknown SKU, a vague request, a crashing tool, and conflicting specs, and each behaved correctly and appeared in its own trace. ## What we learned
we learned about teamwork, delegating the work.
we learned about the supply chain model, industry analysis
Observability should be built in from the first run. Seeing each step and each blocked line in Prism made debugging faster and gave us evidence for our claims. ## What's next for QuotePilot AI
Replace the sample catalog and benchmarks with a supplier’s real data and price feeds. All data in this prototype is synthetic and labeled as such.
Add the rest of a real quote: lead time, availability, quantity breaks, freight, and alloy surcharges.
Suggest substitutes when the exact item is missing, such as a close alternative the buyer can confirm.
Handle drawings and scanned documents, and send unreadable inputs to a human.