05·AI / ML

Infinity Broker

In betaOwner, product, model fine-tune, agent loop, backend, frontend, deployment.

A white-label AI co-pilot for US luxury brokerages: a self-hosted, fine-tuned LLM running a tool-calling agent loop over real Miami-Dade County property records and a brokerage's own CRM inventory and leads. The pitch is the interface, not another data feed, one plain-language layer that reads across the MLS, CRM, and inbox tools a brokerage already pays for.

ResultA live, functional POC over three real public and first-party data sources, hardened by deterministic honesty guardrails (every figure traced to a real tool result, all arithmetic server-side) and a 169-test suite covering every regression found in live stress-testing.

Infinity Broker cover: a browser-framed product hero on a soft blue gradient, headline 'The interface your MLS never built', beside a live chat demo panel with a brokerage team switcher and a broker asking a real Coral Gables question.

Develop

Skills behind it.

Primary discipline plus the support stack.

Skills demonstrated

AI & Machine LearningPrimary
Self-hosted Qwen3 fine-tune driving a multi-tool agent loop (county search, comps, buyer-matching, lead creation, alerts) with per-tool result pinning and a repeat-call loop break for reliability.
Prompt EngineeringSecondary
The honesty layer: system-prompt tool-use contracts plus deterministic overrides that catch fabricated figures, system-prompt leakage, and follow-up chips that contradict the reply, where prompt-tuning alone proved unreliable.
Backend EngineeringSecondary
Serverless tool loop on Vercel, Neon Postgres for persisted leads and hourly alert polling, and integrations to Miami-Dade GIS / Property Appraiser plus a RESO Web API MLS feed (bearer-auth path built, dataset pending).
Product StrategySecondary
Repositioned the product away from 'another data source' (already commoditised by the MLS bundle) toward the orchestration layer, the one conversation across MLS + CRM + inbox.
Frontend EngineeringSupporting
The white-label demo: team-switching chat over real records, a CRM pipeline dashboard, per-listing pages with an honest 'no photography on file' fallback, and a full-screen mobile chat takeover.

Develop · Forks

Decisions on the record.

The few calls worth defending. Each one is a fork; the other branch would have been a different project.

Decision · 01
Why deterministic guardrails, not just a better prompt
Every reliability problem this POC hit, a fabricated square-foot figure, a follow-up chip that re-asks what the reply just answered, the model reciting its own system prompt, came back the moment it depended on the model choosing to behave. The fix that held is a code-level pass that diffs the model's stated numbers against the real tool result and rewrites the reply from real data on a mismatch. Prompting sets the intent; deterministic overrides enforce it.
Decision · 02
Why a self-hosted fine-tune for the POC, Bedrock for production
The self-hosted Qwen3 fine-tune made the POC free to iterate on and fully private, the right call for finding out what the product even is. But its own measured tool-call reliability (~80% ceiling, worse under load) is the evidence for routing production reasoning through Claude on Bedrock instead, with the fine-tune demoted to narrow phase-2 tasks like copy and lead scoring.
Decision · 03
Why 'no photography on file', not stock imagery
County parcel records genuinely carry no listing photos, so the per-listing page shows an honest aerial-and-monogram fallback rather than a decorative stock photo that would imply data the record doesn't have. The same discipline runs through the whole product: a saved sample is labelled a sample, a disconnected capability is declined plainly. When the user is a broker, trust is the product.

Deliver

What shipped.

A live, functional POC over three real public and first-party data sources, hardened by deterministic honesty guardrails (every figure traced to a real tool result, all arithmetic server-side) and a 169-test suite covering every regression found in live stress-testing. MLS asking prices are the one piece still pending a brokerage's RESO credentials.

By the numbers

Automated tests

every live regression pinned

Live data sources

GIS parcels, appraiser, CRM

Warm response

chip suggestions decoupled

Frames