← All work Case study · own product

illeszt — from messy orders to an approved order sheet

AI-powered SaaS for small businesses: it reads the orders that already arrive and lets a human approve instead of retype.

Role
Founder, PM, sole builder
Timeline
Aug–Nov 2026
Status
Private beta, pilot running
Stack
Next.js, Postgres, Claude
Problem

A fresh-produce wholesaler (pilot customer) receives orders as free-text emails and scanned handwritten sheets. Someone retypes them every day.

Customers will not change how they order.

Insight / decision

Don't redesign the customer's ordering habits. Read what already arrives and make a human the approver, not the typist.

Key decisions
Decision 1

OCR bake-off on real sheets

12/12 vs. unusable

Two providers compared on real handwritten order sheets. Claude Vision read the day-column headers in every block; the alternative lost the grid.

Decision 2

Two thin LLM calls

~60 HUF / order

Extract first, then resolve only unresolved rows with a fuzzy top-10. The catalogue stays out of the prompt: roughly $9/month instead of $20/month per the cost model; ~60 HUF per order measured.

Decision 3

Price on the replaced cost

~1.9M HUF / yr

Planned saving at the pilot customer anchors the pricing. Inventory management was dropped from the tiers after the customer interview.

How it flows
  1. Email /
    Scanned sheet
  2. ExtractClaude
  3. Matchlexicon + fuzzy
  4. Human approvalthe gate
  5. Purchasing summary +
    Delivery notes
Evidence
47

user stories

238

commits

135

test files

  • BRD, FS, technical design, pricing model, service design, DPA and terms written
  • LLM regression tests with recorded responses as a PR gate
  • Prompt version logged on every call
Outcome so far

Purchasing summary ready about 1 hour earlier each day at the pilot.

v1 gate in progress: 10 working days of parallel run.

Product
Screenshot: approval screen (to be added)
← All work