Review total
Grand total $52,300 · math mismatch
Line-item sum $48,750 · difference $3,550
Container MSCU1234565 · check digit valid
Auto-pass
Consignee matched against shipment packet
Vessel / voyage consistent across pages
ROUTING DECISION
2 safe fields auto-pass. 1 risky total stops for review with evidence.
The real risk
The real risk isn't your AI bill. It's not know hwo to iterpret.
You're charging a flat price for a cost that swings 100× per call. Without per-call economics, every pricing and model decision is a guess.
Tier that loses money
A mismatched total, missing party, or invalid container value stalls the shipment.
Power user eating the margin
Wrong quantities, prices, totals, or currencies become downstream exceptions.
Model swap you can't justify
A smudged member ID or missing intake field can quietly break the workflow.
Board asks "what's your gross margin?"
Ops teams find the error later, then restart the same document loop.
Provider price hike, invisible until the invoice
The document did not add up, but the system accepted the value anyway.
Pricing set by gut, not data
The customer sees the exception before your team catches the bad field.
How it works
From a single line of code to a board-ready margin report
From a single line of code to a board-ready margin report — without a database, a dashboard server, or a vendor holding your data.
01
Record
Add ledger.record({…}) after each AI call. Fire-and-forget, buffered, never blocks your user.
02
Stamp
Each event is cost-stamped (multi-modal: tokens, images, characters) and provenance-tagged, then written to your object storage.
03
Query
DuckDB reads your Parquet on demand — no server, no database to run. Cost reports, simulations, forecasts
04
Simulate
You change pricing or models in your own product. The ledger informs every call; it never makes one for you.
Send one painful document workflow.
We’ll run the full end-to-end process on your real documents and show what can safely auto-pass, what needs review, and what clean output would look like.
Beyond AI observability
AI observability tools tell you a call happened. They were never built to tell you if you made money on it.
OCR asks: "Show me the trace."
"Show me what a model swap saves before I make it."
Our System: Is this safe to use?
"What margin did we make, by tier and user?"
"Show me what a model swap saves before I make it."
Where & When We Need AI
We don't let an AI do your accounting.
The money math is deterministic, testable, and replayable. AI only writes the report — never the numbers.
Loop 01 · Reduce manual review
Exact, not estimated.
Costs stored as integer micro-dollars. No floating-point drift across millions of events.
Loop 02 · Catch costly mismatches
Deterministic core.
Cost, compaction, and margin math are plain code — no LLM in the path. Same inputs, same answer, every time.
Loop 03 · Improve over time
Isolated by project.
Dev experiments physically can't contaminate your prod cost report. Each environment is its own event stream.
Non-AI layer · Keep trust high
Humans accept, never auto-applied.
Skills propose a price or a model swap. You decide. By design.
Proof example
Your dashboard says 400 Pro users. The ledger says 38 of them are unprofitable.
Pro is priced at $20/month. Most users cost you $7 in AI — a healthy 65% margin. But 38 power users run heavy image + agent workloads that cost $31 each. You're paying $11 a month for the privilege of keeping them. No billing system will tell you this. Your ledger surfaces it on day one — and simulates the fix before you ship it.
Printed total
$52,300
Line item sum
$48,750
Difference
$3,550
Container number
Invoice total
Best-fit customers
AI SaaS builders who starts worrying about their AI Margin.
High-volume document operations
ou charge a flat subscription but your per-user AI cost swings wildly
You mix text, image, and embeddings across providers
You're about to raise prices and you're guessing
Customs brokerage validation
Common checks: invoice total vs line-item sum, container number check digit, required fields, date validity, party names, shipment references, currency normalization, and duplicate or conflicting fields.
Pilot Engagement
Full economics. Controlled scope.
Your data never leaves your storage. The library installs and uninstalls cleanly — no lock-in, no central account.
Pilot I
Cost Visibility
For one painful document type and a small controlled test batch.
✓
1 live product, 1 environment
✓
SDK wired into your real LLM calls
✓
Events flowing to your own storage (R2/S3 bucket)
✓
First 30-day cost report: spend by feature + model
✓
First cost report in under a week
✓
Margin verdict on your current pricing
Start Pilot I
Pilot II
Margin & Forecast
For a stronger production signal across two related document types.
✓
2+ products or multi-environment, federated reporting
✓
Pricing-tier design from real usage histograms
✓
What-if model-swap simulations on your history
✓
Scheduled margin-monitor alerts (cron / CI)
✓
Board-ready 12-month cost + margin forecast
✓
model what-if simulations
✓
Production handoff + ops runbook
Start Pilot II
