LLMs for Arbitrage in New-Build Real Estate

Grigory Zlotin23 May 2026

Problem

Avito sends potential buyers (leads) to new-build developers, who pay for each good lead. If a lead misses an agreed criterion (location, price, finishing and so on), the developer files a complaint, and a manager decides whether to accept it. Developers dispute 10–15% of the 150–200 thousand leads transferred each month, and the volume grows. Manual review costs a lot of labor, cannot be extended to preventive checks, and errors cost money (an accepted bad complaint) or developer trust (a wrongly rejected one). External LLMs (such as Claude or GPT) are not allowed for legal reasons.

Example from the article: a manager spends on average 2 minutes per complaint, up to 10 minutes on long dialogues, reading a call transcript against the contract criteria.

Solution in brief

Fine-tune a LoRA adapter on a 7B model to imitate a large model that explains the manager's historical decisions. The small model returns a verdict, reasons from a fixed list, quotes and a justification. First it works as an assistant to managers, then it handles safe scenarios automatically.

Questions this case answers
  • How to automate adjudication of disputed decisions with an in-house small model that must return a verdict, reasons from a fixed list and cited evidence
  • How to distill a large model's explanations into a small model without losing agreement with human decisions
  • How to automate complaint-adjudication decisions with an in-house distilled small LLM that returns a verdict, reasons and cited evidence
  • How to distill a large LLM into a 7B model without losing agreement with human decisions

Tools and methods

  • Avibe 7B + LoRA — the small model for internal deployment
  • Qwen3-235B — generates reasoning traces for training
  • Distillation on reasoning traces — large model to small model
  • Structured JSON answer — verdict, reason, comments, justification
  • PostgreSQL — storage for precomputed recommendations
Continue reading — it's free

Google or a one-time email code. No password.