Programme note / Applied AI

Footing

A white-label AI co-pilot for US luxury brokerages: one plain-language interface over the MLS, CRM and public records.

My role
Owner, product, model fine-tune, agent loop, backend, frontend, deployment.
Origin
Passion project
Built
2026
Status
Archived
Access
No working demo: the hosted backend is offline
Stack
Static HTML / vanilla JS, Vercel serverless functions, Ollama, Qwen3 fine-tune, Neon Postgres, Miami-Dade GIS, RESO Web API
The Footing landing page: a cream nav with the Footing mark and links to the live demo, how it works, coverage, the CRM dashboard and the review record, over a near-black page with three columns headed Where it sits, What it keeps and What it admits, the headline 'we tell you what we don't know.' in very large type, and a strip reading Every figure comes from a named public record, No figure is written by the model.

Chapter 01 of 04

The brief

A white-label AI co-pilot for US luxury brokerages, expanding from a Miami-Dade proof of concept toward a nationwide layer over the MLS, CRM, and public records a brokerage already runs.

Code complete; the hosted demo was taken down in October 2026.

A brokerage's actual stack already covers most of what an agent needs: MLS, CRM, county records, inbox. The friction isn't missing data. It's that each source lives behind its own login and its own mental model, so an agent context-switches all day instead of just asking a question. Another data source would be commoditised the moment it shipped; the MLS bundle already sells that.

Footing (renamed from Infinity Broker 2026-09-09) wraps that answer in a white-label demo a brokerage can put its own name on: the same agent loop, reskinned per team, over the brokerage's own inventory and leads.

Solo build, expanding from a Miami-Dade proof of concept toward a nationwide market. Code complete over three real data sources (county GIS parcels, the Property Appraiser, and a CRM) and was live before the hosted backend went unreachable.

Not another data feed. The brokerage already pays for the data. What it does not have is one place to ask a question in plain language and get all three to answer.

Chapter 02 of 04

The problem

Give a US luxury brokerage one plain-language interface over the MLS, CRM, and public records it already pays for, instead of selling another data feed.

Chapter 03 of 04

The build

AI & Machine Learning
Self-hosted Qwen3 fine-tune driving the multi-tool agent loop, kept reliable with per-tool result pinning and a loop break.
Prompt Engineering
Honesty as a layered contract: the system prompt sets it, deterministic code enforces it.
Backend Engineering
Serverless tool loop on Vercel; Neon Postgres for leads; county GIS, appraiser, and RESO integrations; a Bedrock backend written but not yet provisioned.
Product Strategy
Repositioned from another data feed to the orchestration layer, then from a Miami-Dade demo to a nationwide one. The interface is the product.
Frontend Engineering
The white-label demo a brokerage can put its own name on, honest fallbacks included.

Build notes. How it came together and the decisions on the record.

[ Read the full build notes ]

Chapter 04 of 04

What it did

Code complete over three real public and first-party data sources, live in Miami-Dade before the hosted Ollama backend went unreachable.

Deterministic honesty guardrails keep it accurate (every figure traced to a real tool result, all arithmetic server-side), backed by a regression suite pinning every fabrication found in live stress-testing. Nationwide expansion (Florida, North Carolina, Wisconsin importers) is written but not yet loaded. The county sources were queried live against Miami-Dade's own endpoints rather than imported as a snapshot, and the CRM inventory behind the demo is generated sample data, not a brokerage's book.

Everything but MLS asking prices ran on real data: county records, CRM, the agent loop itself, not a mocked dataset standing in for a source that isn't connected. That code is complete; the hosted backend that served it is currently offline. The remaining gap, a brokerage's own RESO credentials for live MLS asking prices, was never about the build itself.

22
Honesty-guardrail testsof 425 total across 17 test files
3
Data sources integratedCounty GIS parcels, the Property Appraiser, a CRM; none reachable while the backend is offline
<2.5s
Warm responsechip suggestions decoupled
  1. 01CRM dashboard, seeded demo data
  2. 02Per-listing page, with the honest no-photo fallback
  3. 03Mobile chat
  4. 04The same product as amber text on a 1983 luggable, because the interface is language
Next on the programmeImmersive 3D / Feature / 2026
A dense white point cloud of an open-plan office floor on black, ceiling grid, structural columns and rows of desks with monitors and chairs, three faint circular markers floating above the desks. The project name Spatial Asset Intelligence is set at the lower left in white, above the line "Scan to safety register".

Spatial Asset Intelligence. Turning one handheld building scan into a fire-safety asset register somebody can check, and deciding which classes were dependable enough to hand over.

[ Read the programme note ]

Contact

Pavlo Puzikov, a black and white studio portrait in a pinstripe suit.

Pavlo Puzikov

Marketing and innovation: immersive 3D, applied AI, generative content

Write to

pavlopuzikov@gmail.com

Open to
Freelance projects, remote. Full-time roles, onsite or hybrid in the EU.
Based in
Bulgaria
Most recent role
Marketing and Innovation Lead, BARNES Dubai, 2026. Product Innovation Researcher there from 2025.
Education
BCom, Marketing and Management, University of Auckland, 2020 to 2023
Languages
English, Russian