TREY MOTSENBOCKER
← All work

ALMAGEST: personal knowledge system

STATUS
LIVE: PERSONAL BUILD
TIMELINE
JUN 2026 – PRESENT
SCALE
202 sources · 5 backend services
ROLE
Architecture · Prompt engineering · Application development · Evaluation & testing · Deployment
ARCHITECTURE
Retrieval-augmented generation over a vector store, with a separate trust-scoring layer that rates each source's support for a claim independently of the answer.
STACK
React · Supabase (pgvector, edge functions) · Claude API · OpenAI embeddings

BUSINESS PROBLEM

Notes, sources, and research accumulate faster than anyone can organize them, and asking an AI about your own material usually means it can't tell a solid source from a weak one.

WHAT I BUILT

A knowledge system with a conversational retrieval interface, a 3D map of how the knowledge connects, and a trust-scoring layer that rates how well each retrieved source can actually support the claim it's being used for.

RESULT

In daily personal use. Development surfaced and fixed a real defect: a source-deduplication bug that let one document be counted several times and inflate a claim's confidence score. The first assertion-based test suite in the project caught it.

MY ROLE

  • Architecture
  • Prompt engineering
  • Application development
  • Evaluation & testing
  • Deployment

HOW IT WORKS

  1. 01Sources are ingested, chunked, and embedded so they can be retrieved by meaning rather than keyword.
  2. 02A question is answered from retrieved passages, with the supporting sources shown alongside the answer.
  3. 03A trust-scoring layer rates how well each retrieved source can support the specific claim it is attached to, rather than scoring the source in the abstract.
  4. 04A 3D map renders the relationships between sources and claims, so gaps and clusters are visible instead of implied.

TECHNICAL IMPLEMENTATION

React front end against Supabase, using pgvector for embedding storage and similarity search and edge functions for the retrieval and scoring services.

Embeddings from OpenAI; answer synthesis and claim evaluation through the Claude API.

Retrieval de-duplicates at the source level before scoring, so repeated passages from one document cannot each contribute to a claim's confidence.

An assertion-based test suite exercises retrieval and scoring against fixed expectations, which is what caught the de-duplication defect.

LIMITATIONS

  • Single-user. There is no sharing, permissioning, or multi-user model.
  • Trust scoring is my own scheme, not a published or externally validated metric.
  • The corpus is 202 sources. Behaviour at an order of magnitude more material is untested.
  • The 3D map is useful for orientation, not analysis; it does not yet answer questions on its own.