Overcoming Loneliness with LLM Dating Assistant
summary
Social Discovery Group builds a dating copilot that generates two suggested replies in-app, trains models by fine‑tuning frontier APIs, and deploys across a 60+ app portfolio reaching users in 150 countries. Evaluation uses automated benchmarks (tone, empathy, relevancy) with LLM-as-oracle and OpenAI structured outputs, offline human assessors, then side‑by‑side online A/B tests; iterative fine‑tuning raises engagement in production and reduces inference cost versus raw API usage.
AI-written summary · about
chapters (12)
- 0:00LLM dating assistant intro
- 1:11Why build the copilot
- 2:34Train your own LLM?
- 4:38Model improvement cycle
- 5:39Why measure models
- 7:05Three quality methods
- 9:57Benchmarking pros and cons
- 11:02Online metrics with A/B tests
- 13:40Human evaluation offline
- 16:10Online vs offline evaluation
- 17:35Final model pipeline
- 19:22Q&A
related talks
22mFrom a simple chatbot to a fully programmable dating coach
24mRubyLLM: Making AI Development Beautiful Again
29mBeyond Solo: Enhancing LLMs through Collaborative Dialogue with GAN-Inspired Approaches
30mTowards Clinically Grounded LLMs
33mDesigning LLM Applications for Massive Datasets
29mGenerative-AI: Usecase-Specific Evaluation of LLM-powered Applications
30mLLM-Based Recommendation Systems
24mDemocratizing data access through LLM