Synapse
Semantic search and relationship-mapping engine over personal networks. Query your contacts like a database.
Overview
Synapse turns your unstructured social graph into a queryable knowledge base. Instead of scrolling through contacts or LinkedIn connections, you ask natural language questions: "Who do I know that's built fintech products?" or "Founders in my network working on AI infra."
Problem
Your network is one of your most valuable assets, but it's stored as a flat list of names. There's no structure, no searchability, no way to surface relevant connections when you need them. Contact apps give you names and numbers. LinkedIn gives you a feed. Neither gives you insight.
Approach
- Ingestion: Pulls contact data from multiple sources, normalizes it into a unified schema
- Embedding: Each contact is embedded with metadata (role, company, interactions, shared context) using OpenAI embeddings
- Semantic search: Natural language queries are embedded and matched against the contact corpus using cosine similarity
- Relevance scoring: Contacts are ranked by a composite score combining semantic similarity, recency of interaction, and relationship strength signals
- Interface: Next.js frontend with a search-first UI backed by Ollama for local query understanding
Key Decisions
- Composite scoring over pure semantic search: Raw embedding similarity returns too many false positives. Adding interaction recency and relationship signals dramatically improves result quality.
- Ollama for query parsing: Before hitting the embedding search, the query is parsed locally to extract intent and constraints. This prevents the "everything is similar" problem with naive embedding search.
- Rapid API for data enrichment: Augmenting sparse contact data with public information improves embedding quality and search accuracy.
Challenges
The cold start problem: most people's contact data is sparse. A name and email isn't enough to build a useful embedding. The system needs to infer context from communication patterns, shared events, and public information to build meaningful representations.
Privacy was also a constant constraint. The system stores and processes personal relationship data—every design decision had to account for data minimization and user control over what gets indexed.
Outcome
Turns a passive contact list into an active intelligence layer. The system surfaces connections you didn't know you had and makes introductions that would otherwise require hours of manual network scanning.