Ssokdak-Ssokdak
A digital therapeutic that treats early-stage depression through emotion journaling
Context
A digital therapeutic built on emotion journaling. I led the product engineering team while owning the two pieces at its core — emotion extraction and partner matching.
Early intervention does not follow from entries piling up — it has to lead to a next step. But a journal entry is free-form text with no structured fields, so on its own it cannot serve as recommender input. It needed two stages: pull the emotional signal out of the body text, then compose that signal with other factors to match a person.
How I solved it
- Fixed the emotion label set and made the LLM choose inside it
- the point here is turning free-form text, which names the same feeling differently every time, into a value the recommender can read as is
- Extracted those labels from the journal body and accumulated them per user
- Fed the extracted labels, demographics, and keywords into a neural CF model as factors to match a user with someone to talk to
- learned from interactions rather than paired by rules
- Ran extraction and matching as batch jobs, off the path that saves a journal entry
- Checked extraction output and match quality against metrics
Result
- Launched Ssokdak-Ssokdak v1.0
- The TIPS selection secured ₩500M
- the emotion extraction and the recommender were the core technology behind it