Loopkeeper
A customer-service memory system built on a graph database and an LLM that connects to WhatsApp groups, triaging and classifying messages and linking them to tickets and school deployment information so that issues are not forgotten.
Timeline (last 90 days)
Loopkeeper is a support system developed by Ednovax, launched as Ednovax Support, attached to schools' WhatsApp groups, and it processes messages automatically.
Loopkeeper uses the Claude Haiku model for message triage, information extraction, and question answering, and can recognize Nigerian English and mild Pidgin.
Loopkeeper uses the local embedding model all-MiniLM-L6-v2 (384 dimensions) to vectorize every WhatsApp message, requiring no additional API calls.
Loopkeeper's graph database is divided into three tiers: lk_hq stores Ednovax's own data, lk_client_<id> is one graph per school, and lk_hub is the staff central support group.
The Loopkeeper system contains 182 named Cypher queries, used to implement functions such as semantic search, automatic assignment, and change impact analysis.
Loopkeeper uses the graph algorithm algo.WCC to cluster and deduplicate tickets connected via SIMILAR_TO relationships, and the model only adjudicates whether they are truly identical.
In HQ, Loopkeeper creates an abstract Signal for each ticket (feature, severity, status, and embedding vector, with school information erased), and similar Signals are clustered into Patterns to discover cross-school issues; the build fails if message text or contact names appear in HQ.
When indexing code repositories, Loopkeeper converts import relationships into IMPORTS edges and CODEOWNERS into OWNS edges, never stores the source code itself, and reads the closest files from GitHub in real time when a ticket is opened.
All Facts (8)
Loopkeeper is a support system developed by Ednovax, launched as Ednovax Support, attached to schools' WhatsApp groups, and it processes messages automatically.
50%UnverifiedLoopkeeper uses the Claude Haiku model for message triage, information extraction, and question answering, and can recognize Nigerian English and mild Pidgin.
50%UnverifiedLoopkeeper uses the local embedding model all-MiniLM-L6-v2 (384 dimensions) to vectorize every WhatsApp message, requiring no additional API calls.
50%UnverifiedLoopkeeper's graph database is divided into three tiers: lk_hq stores Ednovax's own data, lk_client_<id> is one graph per school, and lk_hub is the staff central support group.
50%UnverifiedThe Loopkeeper system contains 182 named Cypher queries, used to implement functions such as semantic search, automatic assignment, and change impact analysis.
50%UnverifiedLoopkeeper uses the graph algorithm algo.WCC to cluster and deduplicate tickets connected via SIMILAR_TO relationships, and the model only adjudicates whether they are truly identical.
50%UnverifiedIn HQ, Loopkeeper creates an abstract Signal for each ticket (feature, severity, status, and embedding vector, with school information erased), and similar Signals are clustered into Patterns to discover cross-school issues; the build fails if message text or contact names appear in HQ.
50%UnverifiedWhen indexing code repositories, Loopkeeper converts import relationships into IMPORTS edges and CODEOWNERS into OWNS edges, never stores the source code itself, and reads the closest files from GitHub in real time when a ticket is opened.
50%