Myngly
An AI matching engine that explains every match it makes.
The problem
Myngly matches people who are at the same event. Its recommendations ranked people mainly by distance and app activity, so you saw whoever was nearby and busy, not whoever was worth meeting, and nothing told you why. Vicentius Apps wanted matches based on what people actually offer and what they are looking for, each with a reason attached, without rebuilding an app people were already using.
Constraints
A live product with real users. The AI had to fit into the existing platform, not replace it. AI cost per match had to stay small. People can switch AI matching off. Every match records how it was produced, so it can be traced.
Architecture
- 01 PROFILE What each person offers and is looking for
- 02 EMBEDDINGS Dual embeddings
- 03 CANDIDATE SEARCH Proprietary shortlist
- 04 SCORING Compatibility, both ways
- 05 LLM EVALUATION Grounding and explanations
- 06 VALIDATION Blocked and declined people removed
- 07 STORED MATCH WITH EXPLANATION Saved with its explanation
- CLIENT
- Mobile app
- API
- Existing backend
- MODELS + STORAGE
- Language model · embedding model · vector index
FIG. 1 Matching pipeline. Teal marks where the LLM evaluates each candidate with grounding and writes the explanation that is stored with the match.
What we built
AI inside
Each profile is read for two things: what the person offers and what they are looking for. A match has to make sense both ways. Myngly’s own matching algorithms build a shortlist, then a language model checks each pair against what both people actually shared and explains, in plain words, why they should meet.
End to end
- Product concept B7DEV
- UI and UX design B7DEV
- iOS and Android app B7DEV
- Server side B7DEV
- DevOps B7DEV
- AI matching engine B7DEV
- Optimisation to industry standards B7DEV
Results
Built and run by b7dev since 2022, from first sketch to live app. Every AI match arrives with a reason the person can read. The AI works alongside the matching people already trusted, so nothing that worked was lost. Recommendations stay fast, and every match can be traced, so quality improves with evidence, not guesswork.
Building something similar?
20 minutes, no slides, an honest opinion even if you don't hire us.