Confidential · Internal
Confidential · Internal
An AI-augmented screening platform for the tax-lien property engagement. The system decides the volume automatically; FBSPL's bench handles only the judgment calls. This is the FBSPL-side view — implementation, engagement model, and the total economics of the engagement.
Each property runs through one automated pipeline. Every stage uses free, open data where possible and local AI inference, so the only paid input is a single set of street-view images. Here is the flow end to end — then the technology behind each step.
Resolve the coordinates to a real parcel, its boundary and footprint, and a postal address. This is also where institutional sites (schools, churches, hospitals) are caught from public records and rejected before any AI runs.
Fetch and stitch high-zoom aerial tiles, crop to the parcel, and overlay the exact boundary so the AI judges the right patch of land — not the neighbour's.
A free metadata call finds the camera's real road position, then the building's footprint is used to aim three facade-aligned images at the property face. This is the only paid input in the pipeline.
A vision model reads all four images in a single call and returns a structured verdict — property type, physical condition, and a confidence score — as clean JSON. It runs on our own GPU, so there is no per-call cloud-AI cost.
Three layers in priority order convert evidence into a verdict: public-record hard rejects, the fund's own approve/reject rules applied to the AI's observations, then a confidence gate that routes anything uncertain to a human.
Each property gets a saved record — annotated images, reasoning and confidence — plus a batch result. Delivered as an Excel export today, or pushed straight into the client's CRM if integrated.
The same logic the client's analysts apply by hand, encoded once. The human only enters where the machine is genuinely unsure.
Competitors can sell screening software. FBSPL can sell the outcome: AI volume, human judgment, and a defensible record, delivered by a team that already exists. That combination is the moat. Below, the two engagement models — and our recommendation.
Outcome over software, human judgment built in at near-zero marginal cost from the existing bench, near-zero AI cost from running inference on our own GPUs, and an audit-grade record behind every decision — all on a platform that's built once and reused across funds.
Lead with the managed solution as the primary offer — it is where FBSPL's advantage is greatest — and keep SaaS as a secondary tier for clients who insist on running it themselves.
Each of these changes either the architecture, the cost, or the contract — so they're settled with the client up front.
Managed solution or self-serve SaaS? This sets the whole commercial and operational shape.
Do we host on FBSPL's servers (the managed default), or does the client require it on their own infrastructure for data residency? This is a direct question for the client.
Do they have a paid key to share — so imagery bills to their account — or does FBSPL provision and bill it back? Drives cost ownership.
Stay on the Excel handoff (the MVP), or build direct CRM endpoints plus a review queue inside their system?
Confirm the live approve / reject criteria so the rules and the model target match real policy.
Confirm ~100,000 once a year vs. a variable or more frequent flow — it changes the infrastructure plan.
The full delivery cost of a 100,000-property run — GPU compute, the hosting server, and street-view imagery. The imagery is the swing factor; everything else is small and fixed.
| Setup | GPU | Props / hour | 100K duration | Compute cost |
|---|---|---|---|---|
| 1 worker | RTX 3060 Ti | ~275 | ~15 days | ~$10 elec |
| 3 workers | RTX 3090 (×2 Ollama) | ~1,500 | ~2.8 days | ~$95 RunPod |
| 6 workers | RTX 3080 | ~1,650 | ~2.5 days | ~$151 RunPod |
Vision inference is the bottleneck — it runs one property at a time, so adding GPU workers (rented on RunPod) is what compresses the run from ~15 days to ~2.5. Imagery cost is constant regardless of worker count.
| Per annual run (100,000 properties) | Client's Google key | FBSPL's Google key |
|---|---|---|
| Google street-view imagery | $0 client | ~$1,300 |
| RunPod GPU compute (6× RTX 3080, ~2.5 days) | ~$151 | ~$151 |
| Hosting server (app + review queue) | ~$50 | ~$50 |
| Storage | ~$5 | ~$5 |
| Total cost to FBSPL per run | ~$206 | ~$1,500 |
Imagery is the swing cost. Billing all 100K at list rate is roughly ~$2,000; the realized figure is lower after Google's free monthly allowance and the properties auto-rejected from public records (which skip street view at no charge). The exact number is confirmed at calibration against the live key. Sources: Google Maps pricing list · Street View Static usage & billing · cost calculator.
| Build scope | Effort |
|---|---|
| Harden + productionise core | ~15–25 dev-days |
| CRM integration (if chosen) | +10–20 dev-days |
Built on an existing working pipeline — "finish and harden," not greenfield. Paid once, then amortised across every future run.
The same platform serves additional funds at near-zero incremental build cost — turning one engagement into a repeatable service line.
The ~10–20% of properties routed to human review are handled by FBSPL's existing testing bench at near-zero marginal cost, which is why no labour line appears in the run costs above. A dedicated, ring-fenced review team, if the client wants one, is scoped and priced as a separate managed-service line.
Before the first production run, a one-time calibration pilot runs on a past batch the client has already screened. We compare the AI's verdicts to the human outcome, measure accuracy, and fine-tune the model and the confidence threshold. It is scoped and priced separately from the recurring run, and it is where the final imagery cost is confirmed against the live key.
Cleared in ~2.5 days at ~$1,500 all-in, with the review labour already on the bench — then repackaged and run for other funds on the same platform. The build is paid once; everything after is leverage.