Zero to operator · 8 chapters
Build a GBrain-Style Knowledge Engine
Build a personal answer system that returns proof, names what it does not know, and improves without rewriting your beliefs in secret.
A practical guide inspired by Garry Tan's GBrain product ideas and the answer-plus-proof-plus-gaps pattern. You will model a personal knowledge graph, attach claims to sources, build hybrid retrieval, add safe consolidation, enforce user boundaries, and evaluate whether the system deserves trust.
Begin chapter oneDeliberate progression
The curriculum
Foundation chapters give you the map. Builder chapters teach you to make controlled changes. Operator chapters teach permissions, recovery, quality, and proof.
- 01FoundationFrom Chatbot to Answer EngineUnderstand the product shape: answer, proof, gaps, freshness, and a next verification step.27 min lesson + lab
- 02FoundationModel the Personal Knowledge GraphRepresent people, projects, claims, sources, and time with relationships precise enough to reason over.36 min lesson + lab
- 03FoundationEngineer Trust, Provenance, and UncertaintyTurn evidence quality, conflict, freshness, and missing knowledge into visible system state.38 min lesson + lab
- 04BuilderBuild a Self-Wiring Ingestion FlowCapture durable pages and turn explicit links into a typed graph while keeping model suggestions outside canonical truth until reviewed.47 min lesson + lab
- 05BuilderRetrieve, Traverse, and SynthesizeBuild the answer path that combines filters, lexical search, embeddings, graph expansion, and bounded synthesis.51 min lesson + lab
- 06BuilderBuild the Dream and Consolidation CycleLet the system improve in the background while preserving evidence, approvals, and rollback.48 min lesson + lab
- 07OperatorSecure and Operate a Multi-User BrainEnforce isolation, least privilege, local-first options, observability, and recovery before inviting real users.54 min lesson + lab
- 08OperatorEvaluate Trust and Ship the ProductMeasure retrieval, graph value, citation support, gap honesty, freshness, security, and operator usability before calling the brain reliable.58 min lesson + lab
Capstone
Ship an Answer-with-Proof Personal Knowledge Engine
Build a source-grounded engine over a real domain with at least fifty nodes, one hundred claims or edges, twenty sources, and two isolated users or workspaces. It must answer direct and multi-hop questions, expose proof and gaps, consolidate safely, survive provider degradation, and let a human correct the graph.
What you must build
- ◆Stable answer contract containing claims, citations, gaps, freshness, conflicts, and next step.
- ◆Graph schema for entities, claims, typed edges, evidence, scope, and time.
- ◆Source-preserving idempotent ingestion with identity and edge proposals.
- ◆Hybrid retrieval with bounded graph expansion and claim-level citation validation.
- ◆Dream cycle with automatic versus approval-required policy, digest, and rollback.
- ◆Database-enforced user isolation, sensitive-provider rules, and redacted traces.
- ◆Declared degraded modes plus backup, restore, and derived-index rebuild.
- ◆Twenty-question evaluation matrix with no-graph ablation and release gates.
- ◆Human workflows for inspect, dispute, correct, approve, export, and delete.
How you prove it
- End-to-end trace from source capture to answer and original evidence.
- Graph-enabled versus no-graph comparison showing measurable value.
- Unknown and conflict answers with useful verification steps.
- Reviewed dream-cycle patch and no-op rerun.
- Passing cross-user attacks and provider-outage fallback.
- Clean restore and index rebuild.
- Human correction receipt and evidence-backed release decision.
