Predicate Ventures
All projects

Deployedplatform

Atlas

A cloud knowledge graph that gives AI agents memory, with provenance on every fact.

A cloud, RDF-grounded memory substrate. Atlas runs on ECS Fargate in its own AWS account. It ingests documents, mail, calendar and code, and resolves them into typed entities with provenance on every fact. It serves the graph through an HTTP API and MCP tools.

  • RDF-star graph
  • SHACL validated
  • PROV-O provenance
  • MCP + HTTP API
curl https://www.predicate.ventures/openapi/atlas.json
  • Ontologies & knowledge graphs
  • NLP & information extraction
  • Data science & evaluation
  • Cloud infrastructure
test files
419
MCP tools, all read-or-ingest
11
HTTP route modules
36
connector families
6

The problem

Vector search returns passages that sound relevant. It cannot tell you that two documents contradict each other, or where a claim came from. Nor can it tell you that a concept has drifted since last quarter. For work that has to hold up under audit, similarity is not enough.

The approach

Atlas grounds everything in RDF because models and agents are its main readers, which inverts the usual property-graph tradeoff. RDF’s verbosity stops mattering, and its entailment semantics start paying for themselves. Every fact carries provenance, and contradictions go to an adjudication queue instead of being averaged away. The ontology grows through a reviewed pipeline, not a one-off modelling exercise.

How Atlas fits together

  1. Sources

    Documents, mail, calendar, code

  2. Ingest

    Normalize, chunk, embed, extract, resolve entities

  3. Graph

    RDF-star triples, SHACL shapes, PROV-O provenance

  4. Retrieval

    Hybrid BM25 and vector, then reranked

  5. Agents

    Served through an HTTP API and MCP tools

From raw sources to a graph agents can query, with provenance kept at every step.

What Atlas does

  • RDF over vector similarity

    Grounded in RDF, so an agent gets surfaced contradictions and provenance on every fact instead of a similarity score.

  • Provenance on every fact

    Bi-temporal, with PROV-O provenance kept through the whole ingest and retrieval pipeline.

  • Hybrid retrieval

    BM25 and vector search combined, then a planner, reranker and synthesizer stage.

  • Served over HTTP and MCP

    Read through an HTTP API or MCP tools, so any agent framework can plug in.

What’s inside Atlas

  • Ingest: normalize, chunk, embed, spaCy, an LLM pass, entity resolution, writers
  • Retrieval: hybrid BM25 and vector, then planner, reranker, synthesizer
  • Storage: an RDF-star graph on encrypted EFS with hourly S3 snapshots, plus a PostgreSQL 16 vector and full-text tier on Amazon RDS. Bi-temporal, with PROV-O provenance
  • Ontology: SHACL shapes, imports of schema.org, FOAF, SKOS and PROV, OWL 2 RL entailment, SPARQL migrations, and a competency-question suite
  • Connectors: GitHub, Gmail, Calendar, Drive, Pipedrive, Hunter, Linear
  • Intelligence: contradiction detection and entity timelines
  • Runs on ECS Fargate in its own AWS account, with Bedrock serving the LLM, embeddings and rerank
  • Served through an HTTP API and MCP tools, so any agent framework can read it

Applied: the memory layer behind Keel (formerly Cofounder)

Atlas is the memory layer behind Keel, Blake's gated autonomous-coding platform. It runs as its own cloud service, on ECS Fargate inside the operator's AWS account, with no public ingress. Keel's backend reaches it over an internal network path and keeps the bearer token server-side. The Keel app itself cannot reach Atlas directly.

The data stays in that account and is encrypted in each store. The relational tier sits on Amazon RDS, the knowledge graph on EFS, and blob snapshots in S3. Each deployment is single-tenant, run for one operator rather than shared across customers.

Every write Keel sends carries provenance an agent cannot forge, scoped to what its token is allowed to touch. New capabilities plug into the same seams every Atlas deployment uses: 11 MCP tools, 36 HTTP route modules and 6 connector families. Each is added through a documented extension point, never a one-off patch.

Read about Keel

  1. Keel app

    iOS, watchOS and web clients

  2. Keel backend

    Holds the scoped bearer token

  3. Atlas

    In-VPC only, no public ingress

  4. Encrypted storage

    RDS, EFS and S3 in the operator's own AWS account

Every read and write to memory passes through Keel's backend; nothing reaches Atlas directly.

Where Atlas stands today

  • 419 test files, deployed and under active development
  • 11 MCP tools, all read-or-ingest, across 36 HTTP route modules
  • An evaluation suite with gold sets, named ablations and versioned recall-floor baselines committed alongside the code

The Atlas tech stack

  • Python 3.11
  • FastAPI
  • SQLAlchemy 2
  • pyoxigraph
  • rdflib
  • pyshacl
  • pgvector
  • scikit-learn
  • Alembic
  • OpenTelemetry
  • AWS ECS Fargate
  • EFS
  • Amazon RDS PostgreSQL
  • S3
  • Bedrock
  • Titan embeddings
  • Cohere rerank