05·AI / ML
Barnes Dubai LLM
A domain-tuned LLM and agentic broker assistant for Dubai luxury real estate, a QLoRA fine-tune on 15,369 instruction pairs, running locally.
ResultLive in the Innovation Portal playground.
Discover
Why it exists.
Atlas surfaces the data; Barnes Dubai LLM is what reasons over it. A multi-agent research system fine-tuned on real-estate jurisprudence, deployed locally via Ollama with zero per-query cost, capable of producing brokerage-grade analyses across ten-plus sources on demand.
The core decision was to fine-tune rather than wrap a frontier model with prompts. Brokerage-grade output needs a model that understands DLD transaction conventions, France DVF reporting cadence, Singapore URA mapping, none of that lives in a frontier model's pre-training in a useful way. 15,369 instruction pairs, hand-curated from five jurisdictions, taught the base model what BARNES Dubai actually means when it says recent, comparable, or flag for review.
StackPyTorch · QLoRA · Ollama · Python · FastAPI · Tool calling · Multi-agent
Define
The brief.
Decide whether a domain fine-tune beats base + RAG for a terminology-heavy, mixed-language vertical, then ship the answer brokers can use on WhatsApp.
Develop
What it took
Skills behind it.
Primary discipline plus the support stack.
Skills demonstrated
- AI & Machine LearningPrimary
QLoRA domain fine-tune + a tool-calling agent loop (getCommunityInventory, comp-search, valuation) with audit-logged human-in-the-loop above the confidence threshold.
- Prompt EngineeringSecondary
Prompt + tool-schema design for the agent loop, system prompts, function-calling contracts, a confidence-gated escalation prompt, and an eval harness over real broker queries.
- Backend EngineeringSecondary
FastAPI inference service with broker_id-keyed conversation memory; Ollama deployment; WhatsApp surface in front of the playground.
- Data PipelinesSecondary
15,369 analyst-reviewed instruction pairs across five national registries, plus live Atlas connectors feeding tool calls.
- ResearchSupporting
Training-run analysis and dataset assembly that targeted brokerage-grade jurisprudence.
Develop · Build
How it came together.
- 01
QLoRA on a domain-specific corpus, not RLHF
QLoRA over a base instruct model was the right tier, full fine-tune was overkill, prompt engineering was undertrained. The instruction pairs were sourced from Dubai Land Department, UK Land Registry, France DVF, Singapore URA, and the World Bank; each pair was reviewed by an analyst so the model learns the team's judgment, not just the public registry's wording.
AI & Machine Learning - 02
Local deployment via Ollama, not a hosted endpoint
Per-query cost on a hosted frontier API would have foreclosed the on-demand research workflow before it shipped. Local Ollama deployment on a single workstation drives the cost to electricity, runs at conversational latency, and keeps all transactional data inside the org's network, non-negotiable for a brokerage handling proprietary listings.
Backend Engineering - 03
A guarded agent loop, not chat completion
The multi-agent wrapper treats each research run as a guarded loop: reason over the request, propose an action, then either commit it above a confidence threshold or escalate to a human below one. Every accepted and rejected action is logged, and that log feeds the next training pass.
AI & Machine Learning - 04
FastAPI in front so other tools can call the agents
Wrapped Victor behind a FastAPI service so Atlas, the Innovation Portal, and ad-hoc Python scripts all hit the same agents through one API. Means the multi-agent system is a shared brain across the BARNES Dubai stack, not a one-off chat surface.
Backend Engineering
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 QLoRA fine-tune, not base + RAG
Base + RAG handles open-domain queries well, but Dubai luxury real-estate is terminology-heavy, mixed-language, and full of low-frequency entities (off-plan towers, niche communities, broker shorthand). The QLoRA pass lifted named-entity recall and got the model speaking the vertical's actual dialect, the place RAG most often hallucinates is exactly where this matters.
Decision · 02
Why a three-mode picker, not one chat box
The first version was a single conversational surface and it conflated three jobs: market analysis (Barnes Dubai LLM), tool-orchestrated workflows (Victor), and marketing-asset generation (Design). Splitting them into a mode picker made the right model + prompt + tool-bundle visible to the user up front, instead of relying on the LLM to figure out intent. Users now pick the task; the model does the task.
Decision · 03
Why Cloudflare Workers AI for vision, not local
A broker shoots a listing on a mid-range laptop with no GPU behind it. What dominates the prompt-eval pass is vision-token tokenisation, not the text the model writes back, so handing the image to a server GPU beats grinding it through the client CPU. Cloudflare Workers AI runs llava on GPU with a 10-15s warm round-trip, and the free tier's 10k requests a day sits well past what a broker desk generates. Only the photo bytes leave the broker's device.
Deliver
What shipped.
Live in the Innovation Portal playground. A broker asks Compare Business Bay vs JVC for buy-to-let yield and gets a tool-resolved verdict in seconds (medians, gross yields, tenant pool, a one-paragraph recommendation), the brief that used to take an analyst an afternoon.
In production with the BARNES Dubai research team. Ten-plus ingest sources, transaction registries, listing portals, social, news, internal CRM, flow through Victor and produce ranked opportunities on demand. The audit trail closes the loop: the next training pass learns what the team chose to act on.
By the numbers
Instruction pairs
QLoRA fine-tune corpus
National registries
DLD, UK, France, Singapore, World Bank
Research origin
The paper-to-product thread this project sits inside. The full chain (paper, benchmark, ship) lives on the research index.
- Domain LLM fine-tuneSee thread
Parameter-efficient adaptation of a base model to Dubai luxury real-estate.
Frames
Selected from the work.
1 frame
01AI Assistant