Authorized records
Contracts, amendments, invoices, payments and credits. Immutable originals and access controls.
03 / System design
From payment records
to a defensible case.
A proposed production architecture for one recovery workflow. Models interpret documents, reproducible rules calculate amounts, and specialists approve claims.
Operating loop
Eight connected stages.
Normalize & reconcile
Stable identifiers, effective dates, amounts, coverage checks and control totals.
Exception rules
Versioned SQL/Python calculations, overlap exclusions and candidate ranking.
AI case preparation
Extract terms, retrieve clauses and draft explanations with exact document references.
Specialist review
Human acceptance, rejection reasons, missing records and a second-review sample.
Partner collection
Authorized supplier contact, disputes, appeals and collection decisions.
Outcome ledger
Posted cash, realized credits, reversals, partner share and Now City fees.
Evaluation & learning
Held-out cases, reviewer time, costs and measured improvements to rules.
Keep the project hub for research, decisions and progress. Approve the transaction environment with each client before ingesting financial records. This static website serves planning content and synthetic fixtures.
Reference implementation
Start with a small, observable stack.
| System | Pilot implementation | What to record |
|---|---|---|
| Website & deployment | Astro static build → Cloudflare Pages → GitHub builds | Commit, build result and deployment URL |
| Raw documents | Approved encrypted object storage, separate client access | Hash, version, origin, access and retention |
| Transactions & cases | Relational database; SQLite for synthetic local work, managed Postgres candidate for production | Contract versions, payments, credits and case history |
| Extraction & retrieval | Text parsing, OCR where needed, full-text search; semantic retrieval after evaluation | Page/clause references and extraction confidence |
| Rules & orchestration | Versioned calculations, scheduled jobs, retries and exception queue | Rule version, input snapshot and reproducible result |
| Model service | Schema-constrained extraction and packet drafting with selected API models | Model/prompt version, actual billed tokens and cost |
| Review & finance | Case review UI, specialist decisions and reconciliation ledger | Review time, accepted value, cash and reversals |
| Security & operations | Identity, least privilege, audit logs, backups, incident process | Client approvals and tested recovery procedures |
Monthly planning model / USD
Token burn is one line in the budget.
Change volume, model rates and labor assumptions. The default represents an ongoing 10,000-document evaluation queue, with every candidate receiving specialist review.
…
Total modeled monthly operations
…
Model tokens, including overhead
…
Candidate packets for review
…
All modeled cost / source document
Editable assumptions
| Monthly cost | Estimate |
|---|
What the model assumes
Every document gets one extraction pass; candidates get one additional preparation pass. Token overhead applies to both. Output assumptions must include billed reasoning tokens where applicable. Repeated contract extraction should be cached or reused after correctness is established.
OCR defaults to two pages per document at an assumed $0.01/page. Set it to zero for already-extracted text. Infrastructure combines storage, database, jobs, logs, backups and site hosting as a planning allowance.
What still needs a quote
Labor, infrastructure, OCR and the fractional security/counsel reserve are assumptions. Initial integration and build work, founder pay, sales, insurance, taxes, enterprise compliance and unusual legal work are additional. Partner fees and working capital require a separate commercial/cash-flow model.
Capacity check: 300 candidates × 15 minutes = 75 review hours per month at the defaults. Fixed engineering and specialist allowances dominate at small volumes. Accepted claims and collected cash remain separate from these throughput costs.
Rate basis and sources
Checked 4 November 2026. Illustrative standard API rates: GPT-5.4 mini at $0.75 input / $4.50 output per million tokens; GPT-5.4 at $2.50 / $15.00. These are cost references, subject to model evaluation and client approval. Cached-input discounts, batch discounts, tool-call charges, residency uplifts and image-token variability are excluded. Replace rates before budgeting a live pilot.
Controls & evaluation
Make every case reproducible.
Trace the amount
Link each calculation to the correct contract version, invoice, payment and credit. Use decimal currency arithmetic and preserve the source record.
Measure incrementality
Freeze the incumbent baseline, exclude previously worked claims and keep closed outcomes out of packet preparation.
Follow the outcome
Record rejection reasons, disputes, review time, posted amounts and reversals. Improve rules using adjudicated outcomes.
Governance reference: NIST AI Risk Management Framework, including human responsibility, evaluation and traceability.