Airon
AI-Native POS & Autonomous Business Intelligence
An Astheron product for independent retail owners.
3×Faster backend I/O

Case Study
The problem
Cashiers retype every supplier invoice by hand — slow, error-prone work that leaves stock data perpetually out of date. Owners end up making pricing and restocking decisions on guesswork rather than numbers.
Architecture
Next.js on the Bun runtime, with PostgreSQL as the transactional source of truth and Redis holding the aggregates that are read most often. Two AI paths run separately: a VLM handles multimodal invoice scanning at input time, while an LLM reads the cached transaction trends to compose analytics summaries — so heavy inference never blocks the cashier flow.
The hard part
Supplier receipts have no standard format: layout, discount placement and item naming differ per vendor. Line-by-line OCR breaks on this. A VLM is used precisely because it reads layout visually — understanding that a figure is a price because of where it sits, not because it matched a pattern.
Outcome
Invoice entry moved from manual typing to instant scanning, and owners receive daily summaries, low-stock projections and pricing suggestions without requesting a report. The Bun runtime keeps backend I/O handling up to 3× faster than conventional Node.
System Architecture
Application
Next.js · Bun
Application layer and the till flow, running on the Bun runtime.
Data
PostgreSQL
Source of truth for transactions.
Redis
Holds the most frequently read aggregates.
AI paths
VLMAI
Multimodal invoice scanning, at input time.
LLMAI
Reads already-cached transaction trends to build analytic summaries.
Why it is shaped this way
The two AI paths run separately, so heavy inference never blocks the till.
Engineering Decisions
Intelligent Input via VLM
Replaces manual cashier data entry with a Vision Language Model that scans invoices and receipts multimodally, extracting item metadata, pricing and discounts instantly.
Autonomous Business Analyst
An LLM reads transaction trends from PostgreSQL cached in Redis, producing daily analytics digests, low-stock projections and pricing strategy suggestions for the owner.
High-Performance Runtime
Adopted the Bun runtime for backend I/O handling up to 3x faster than conventional Node.