DEV Community: Alfred M

Per-Customer Cost Tracking for LLM SaaS: Know Your Margins Before They Disappear

Alfred M — Mon, 04 May 2026 19:14:15 +0000

Per-Customer Cost Tracking for LLM SaaS: Know Your Margins Before They Disappear

If you are building a SaaS product on top of LLMs or third-party APIs, your costs are not fixed. What you might not know is which customers are responsible for the expensive ones.

The Problem

Most small teams have a rough idea of their total API costs. Fewer have visibility into per-customer margins. Your API provider dashboard and Stripe do not talk to each other, and neither answers the question you actually need: what did it cost me to serve customer X this billing period, by feature?

Without that number, usage-based pricing is guesswork.

Why Existing Tools Fall Short

OpenMeter and Lago are solid open-source metered billing tools, but they are designed to bill outward to customers and require infrastructure like Kafka and ClickHouse that is hard to justify for a small team.

Stripe tells you revenue per customer. It cannot tell you margin per customer.

What Per-Customer Attribution Looks Like

Every API call is tagged with a customer ID and a feature name
You define cost rules per feature, for example 0.01 USD per 1000 tokens for GPT-4o
Costs accumulate in real time giving you live margin per customer and per feature
Alerts fire when a customer crosses a cost threshold before the billing cycle ends

UsageMeter is built around this model. A REST API and TypeScript SDK, a few hours to integrate, no infrastructure to babysit. 29 GBP per month.

If you have never verified your pricing actually covers your costs per customer, it is worth finding out before you subsidise someone for six months without realising it.

How to Convert PDF Invoices to Excel Automatically

Alfred M — Mon, 04 May 2026 19:05:57 +0000

How to Convert PDF Invoices to Excel Automatically

If you run a small business and handle your own bookkeeping, there is a good chance that somewhere in your monthly routine you are manually copying supplier invoices into a spreadsheet. Supplier name. Date. Line items. Tax. Total. Next invoice.

This is one of those tasks that should have been automated years ago.

Why It Is Harder Than It Looks

At ten invoices a month, manual entry is annoying but manageable. At fifty or a hundred, it consumes real time and introduces errors. A transposed total or missed line item creates headaches when your books do not reconcile at quarter end.

The core problem is that supplier invoices are unstructured. Every supplier has their own layout and field labels. That inconsistency is what makes automated extraction genuinely difficult.

The Options (and Why Most Fall Short)

Basic OCR tools return raw text. You still have to parse the fields yourself.

Enterprise document processing platforms handle structured extraction but are built for large organisations with dedicated teams and budgets to match.

Neither makes sense for a sole trader or small business.

What Actually Works

Drag in a batch of PDF or image invoices. AI extracts the structured fields: supplier, date, line items, tax, totals. Export a clean Excel or CSV in one click.

ParsePoint does this with a column template system to match your own spreadsheet layout, and a supplier memory feature that learns each vendor format over time. 19 GBP per month with a free tier. Built for sole traders, freelancers, and small ecommerce businesses.

Claude Code Approval Gates: Keep a Human in the Loop When AI Agents Run Shell Commands

Alfred M — Mon, 04 May 2026 19:05:21 +0000

Claude Code Approval Gates: Keep a Human in the Loop

AI coding agents like Claude Code, Codex CLI, and Cursor are becoming standard in engineering workflows — writing migrations, cleaning directories, running database operations, often with minimal oversight.

That is mostly fine. Until it is not.

The Problem: Confidence Without Brakes

The risk is not malice, it is confidence. Claude Code suggests DROP TABLE users in the same tone it uses to rename a variable. If you are not watching when it runs, you find out later.

Most teams have no formal process for approving destructive commands — they rely on whoever is at the keyboard to catch problems in real time.

Which Commands Need a Pause Button

DROP, TRUNCATE, DELETE FROM database operations
rm -rf recursive deletion
git push --force or direct pushes to main
Anything touching .env files or production credentials

How AgentGate Works

A proxy CLI sits between your agent and the shell. Low-risk commands pass through. Anything matching a configurable regex ruleset pauses and routes an approval request to a web dashboard or Slack. One click to approve or reject. Everything logged: exact command, session, approver, timestamp.

Strict rules for prod, relaxed for dev. Role-based team permissions. Works with Claude Code, Codex, Cursor, or anything that shells out.

AgentGate — 49 GBP per month. If your AI coding workflows touch production and your current safety plan is someone was watching, this is for you.