The problem
Wesion connects members by goals, strengths and values. It needed matching, video recommendations and conversation summaries that understood meaning — and the first AI prototype was too expensive to run for every member.
What I built
- 01Designed and built the NestJS back end from scratch on TypeORM and PostgreSQL, deployed on AWS, with a Python Lambda compressing uploaded images and video on S3.
- 02Built 11 AI services — semantic matching, video recommendations, goal suggestions, content analysis and chat summarisation — on OpenAI GPT and LangChain.
- 03Stored each member's strengths, weaknesses, interests, values and keywords in separate Pinecone namespaces, so matches compare like with like.
- 04Measured cost per feature, moved tasks that did not need a reasoning model from o1-mini to gpt-4o-mini, and set token limits of 300–2,000 per use case — 95% cheaper with output quality held.
- 05Added a weekly NestJS cron that summarises each 7-day conversation window into a narrative and action items, with REST endpoints for on-demand digests.
Screens
