Multi-provider AI credits system

As the agency's platform added AI features — chatbots, transcription, scraping, document parsing, video analysis — each capability billed differently: OpenAI charges per token, Firecrawl per scraped page, Unstructured per page/size, Whisper per transcription, Vertex per video second. The business wanted to sell 'AI usage' to clients with a margin, but there was no way to even account for what a user consumed, let alone price it.

• Every provider reports consumption in incompatible units — tokens, pages, seconds, documents — with different pricing tiers that change over time.
• Without metering, a single heavy user could burn unbounded provider spend — unlimited consumption was the default.
• The resale price needed a configurable margin over real provider cost, auditable per operation.

Every AI operation in the platform passes through a metering gateway: the response's usage payload (tokens, pages, seconds — whatever the provider reports) is normalized by a per-provider cost table into credits, the platform's single internal currency. Each conversion applies a configurable margin multiplier, so one credit always maps to a known resale price. Credits are debited atomically per operation against the user's balance in PostgreSQL (with Redis for fast balance checks at request time), and every debit writes a ledger entry — provider, raw usage, rate applied, margin — making consumption auditable end to end. When a balance hits zero, AI features stop gracefully instead of generating unbounded provider bills.

I'd version the rate table from day one — when a provider re-priced, historical ledger entries stayed correct but 'current cost' comparisons got confusing until we pinned rates to time ranges. I'd also add per-operation cost alerts earlier; the first anomalous bill was discovered in a monthly review, not by a monitor.