Operations platform
One platform for sales, collections, inventory, marketing and finance.
Built alone in seven months for a multi-location auto retail and finance group with a $75M+, 4,500-account portfolio. Python, FastAPI and PostgreSQL, with Claude on top.1
- scheduled jobs
- 130+
- automated report deliveries
- 45+
- commits
- 2,700+
- schema migrations
- 250+
AI analytics assistant
Ask the business a question. Get the numbers behind the answer.
The CEO and GM use it from their phones. It picks from 46 tools that query live company data, answers in plain language, and records what every answer cost. Type a question below, or pick one.
Under the hood
Sample data. The demo answers a fixed set of questions.6
Tools, not text-to-SQL
Each tool wraps a report query the business already trusts. The model never writes SQL, so it can't invent a column or run a query that bogs down the database, and its numbers match the morning reports.
Refuses instead of guessing
When no tool fits, it says so and logs the question, so a gap shows up as a request instead of a wrong answer.
Scoped and budgeted
Tools check the user's permissions at call time. Results are capped, tool definitions are cached, and each user has a daily token budget, so the cost per answer stays predictable.
Daily operations and financial reports
Every number, graded against its own history.
Two briefings reach leadership each morning: an operations brief across sales, collections and inventory, and a financial report tracking 22 KPIs. Each card is graded against its own last 30 workdays, and each report opens with an AI-written summary. Tap a card to see how it was graded.
Skip tracing workbench
One click runs the whole search.
When a borrower stops paying and can't be reached, the workbench searches six sources at once, with GLBA and DPPA safeguards. Claude reads what comes back, including people-search reports that run dozens of pages, and writes a dossier: likely locations, ranked, and the next step to take. Each step a collector logs feeds the next recommendation.
Reveals need a permission and are written to the audit log. This demo keeps the number masked.
Sources
AI dossier
Made-up case, people and places6
Grounded AI summaries
The model writes the sentences. SQL supplies the numbers.
Report summaries are written from figures computed in advance. A validator checks every number, trend word and account before anything is sent. Audits with it corrected 11 of 31 trend signs and cut false alerts from 58 to 8.
Sample data6
Inputs, computed in SQL
- cash_collected_wtd
- $412,380
- cash_vs_last_week
- +6.2%
- accounts_45_dpd
- 212
- accounts_45_change
- −4
- broken_promises_7d
- 37
- broken_promises_change
- +9
- call_today
- A-20417, A-19883
Model draft
Validator
Cost controls
AI that costs cents per task.
Every model call logs what it cost, so spend is measured the same way as every other number on the platform.
The cost log caught one model recording $0 and production quietly stuck on an older model. The catalog's evidence check exposed a field the company had been misreading as net loss.
Data reliability
Built on the assumption that data fails quietly.
Health checks, anomaly rules and parity tests against the old spreadsheets. They surfaced a collections report that overstated work six-fold and a queue filter that hid a third of open accounts.
Accounts 45+ days past due that were worked
Old report against real contact records3
- 1,531 accounts, a third of open accounts, sat outside every collections queue. Alerts now catch it.
- New worklists matched the old spreadsheets on 174 of 177 accounts and caught two the sheet missed.
- A mileage-billing check blocked $72,792 in erroneous bills during the dry run.
Speed and accuracy, before and after
On limited on-site hardware
| Measure | Before | After |
|---|---|---|
| Sync deadlock | 12 h | 0.1 s |
| Report page load | 16 s | 1.3 s |
| False AI alerts | 58 | 8 |
| Admin pages | 1x | 11x faster |
- Five hidden failures caught within days, including a pricing outage behind green health checks and a week-long data freeze.
Analytics and models
Models leadership can act on.
A promise-to-pay model built from 227,789 past promises and shipped 5 days after the request. A repossession-value model. Survival and static-pool analysis for policy and lender questions.
Promise-to-pay calibration
Predicted against actual, by decile. Illustration.4
Repossession value, median error
Dollars off the realized value
- Promise-to-pay: within about 6 points in backtesting, with 4x discrimination.
- A 90-day account-flag pilot the GM adopted, with one state as the control group.
Automation
130+ jobs run the week.
Overnight syncs and pipelines, morning report deliveries, and live refreshes through the business day. 45+ reports arrive on their own.
Tech specs
| Scope | Sales, collections, inventory, marketing and finance for a multi-location auto retail and finance group. About 28 staff and executive users. $75M+, 4,500-account portfolio. |
|---|---|
| Built | One person, seven months, with AI coding tools (Claude Code, Codex). 2,700+ commits and 250+ schema migrations. |
| Stack | Python, FastAPI, PostgreSQL, SQLAlchemy and Alembic, APScheduler, Docker, nginx, Cloudflare Tunnel, Playwright. |
| AI | Claude API with tool use, grounded summaries, prompt caching and batch processing. Cost logged on every call through four model upgrades. |
| Data | 130+ scheduled jobs, 45+ automated report deliveries, and the company's first daily history of account data (6.3M snapshots). |
| Integrations | DealerSocket (dealer management), MagiLoop (CRM), PassTime and Ituran GPS (8.5k devices, 5 nightly pipelines), TLOxp, DRN plate readers, TxDMV registrations, J.D. Power and Cox StockWave valuations, Meta and Google Ads, Twilio SMS. |
| Security | Role-based permissions, audit logs of logins, exports and data reveals, encryption of identity data, and 14 automated rules that flag unusual access. |
| Reliability | Health checks on every pipeline, daily anomaly checks on pricing and contracts, and alerts built for data that fails quietly. |