← Work

Sixty Newton Technical Services

An estimator where the AI explains and the code prices

A private quoting app covering all fourteen Sixty Newton work types: material cost from Zoho Books, a pure-TypeScript pricing engine, a Claude assistant that flags over- and under-pricing, and a full quotation PDF with automatic follow-ups.

Role
Head of Technology: product design, pricing engine, Claude integration
Period
2026
Stack
Next.js · Supabase (Postgres, Auth, RLS) · Claude API · @react-pdf/renderer
Abstract render: an obsidian balance of two floating discs, one holding a cyan pellet
  • 14work types priced
  • 12database migrations with RLS
  • immutablequote revisions
  • automaticfollow-up emails · from quotes@60newton.com
  • 4rate provenance levels · books / tds / quote / manual

01Context

Sixty Newton quotes waterproofing, flooring and tiling jobs by the square metre. Estimates were built in spreadsheets from memory of material prices, and follow-ups depended on whoever sent the quote.

02Constraint

A wrong price costs real money on a 6,400 m² job. Any AI in the loop had to be helpful without ever being the thing that does the arithmetic, and every quote revision needed an audit trail.

03What I did

I built the workbench on Next.js and Supabase with row-level security. A pure-TypeScript pricing engine prices stages and product families from rates with provenance: Zoho Books cost, technical datasheet coverage, a previous quote, or a manual entry with a confidence level. Material cost is pulled from Zoho Books. A Claude assistant proposes changes through tool calls and explains each charge, flagging lines that look over- or under-priced; it never computes a total. Quotes are immutable revisions rendered to PDF, synced to Zoho Books by a cron, with automatic follow-up emails and a weekly digest.

04Result

The app covers all fourteen work types and produces a full quotation PDF with an explanation of every charge. Estimators now see where each rate came from and how confident it is before a quote leaves the building.

05What I'd change

Separating proposal (model) from computation (engine) was the right call and made testing trivial. What I underestimated was rate provenance: half the value turned out to be knowing where a number came from, so I'd design that first next time.

06Stack

  • Next.js
  • Supabase (Postgres, Auth, RLS)
  • Claude API
  • @react-pdf/renderer
  • zod
  • Vitest
  • Zoho Books API