What Is It
Enceladus is a production knowledge-governance platform I designed and built from scratch that solves a problem the AI industry is actively wrestling with: how do you get multiple AI agents from different providers to coordinate real work on shared projects without collisions, data loss, or governance gaps — and how do you make the system learn from what they do? The platform manages 20 active production projects, has accumulated 2,789+ governed records across 17 governed lessons, and runs on AWS serverless infrastructure at ~$35/month. I am the sole architect, developer, and operator.
The core innovation is treating project management primitives — Features, Tasks, Issues, and now Lessons — not as database records, but as ontologically defined objects with governed lifecycles, required evidence gates, and deterministic completion contracts. The Lesson Primitive (ENC-FTR-052) gives the platform a first-class governed record type for institutional wisdom: every lesson is constitutionally scored against four philosophical pillars and a vibe board before it can influence governance. Every mutation flows through a single MCP server — 5 code-mode tools delivering an 89% schema reduction — enforcing governance authorization on every write. The result: any AI agent with credentials can safely participate in governed development workflows, no task advances on trust alone, and the system accumulates knowledge at the architectural level.
Architecture & Design Decisions
20 Lambda functions, 8 DynamoDB tables, 2 SQS FIFO queues, Neo4j AuraDB Free (graph-indexed search), CloudFront CDN, API Gateway HTTP v2, Cognito auth with Lambda@Edge
React 19 PWA with TypeScript, Vite, Tailwind CSS, TanStack React Query — mobile-first governance cockpit across all active projects with full primitive surface
Code-mode MCP: 5 governed meta-tools (search, execute, get_compact_context, coordination, connection_health) over Streamable HTTP + OAuth 2.1/PKCE — 89% schema reduction vs. raw mode
5 GitHub Actions workflows, nightly SHA-256 parity audits, secrets guardrail, 13 deployment types with semver changelog; component registry enforces deployment transition arcs at the infrastructure level
Coordination API with dispatch heuristics and graph-indexed tracker search across Claude, OpenAI Codex, and AWS Bedrock; 10 typed relationship edges with weight, confidence, and provenance
Systems Thinking & Key Innovations
Governed Lesson Primitive
The Lesson Primitive (ENC-FTR-052) transforms operational history into institutional knowledge. Every lesson is a first-class governed record type, evidence-gated and append-only, constitutionally scored against four pillars (force/surrender, convergence/play, efficiency/love, intention/flow) and a vibe board before it can propose governance amendments. Cross-project knowledge mining scanned 2,789 records across 20 projects and produced 17 governed lessons — the first time Enceladus examined its own operational history as a knowledge asset.
Exclusive Checkout Service
Prevents agent collisions via atomic task ownership — only the owning session can advance status. Child tasks support parallel dispatch. Commit Approval IDs (CAI) gate code completion; Commit Complete IDs (CCI) are validated in PR bodies by GitHub Actions before merge. A checkout-service assistant subsystem auto-remediates misclassified tasks by relaxing deployment arcs when evidence confirms the mechanism, and can never tighten them.
Evidence-Gated Lifecycle
Task state machine (open → in-progress → coding-complete → committed → pr → merged-main → deploy-success → closed) requires proof at every gate: commit SHAs validated against GitHub API, PR merge timestamps within 60-second tolerance, deployment evidence from GitHub Actions Jobs API with 7 validated fields. The component registry enforces the most-restrictive deployment arc across all components a task touches — infrastructure-level governance agents cannot bypass.
Governance as Architecture
SHA-256 governance hash required on every write mutation. MCP-API boundary policy ensures no tool handler directly accesses DynamoDB business tables, preventing transport-specific behavior drift. Write-source attribution on all mutations enables the audit Lambda to detect anomalies and alert via SNS. A CI-gated governance data dictionary enforces authoritative field semantics at both runtime and deploy time.
Token Economy Design
Every design decision weighs token cost. The code-mode MCP interface delivers an 89% tool-schema reduction. Graph-indexed tracker search collapses 20–50 sequential record lookups into 1–2 Neo4j traversal queries. Context nodes wrap every tracker record in mathematically-scored metadata enabling greedy knapsack context packing within token budgets. Prompt caching (90% discount), strategic model selection, and batch API routing (50% discount) compress operational AI cost by an estimated 60–80% overall.
Event-Driven Pipelines
DynamoDB Streams → EventBridge Pipes → SQS FIFO (natural debounce via 5-min visibility timeout) → Lambda feed publisher → S3/CloudFront invalidation. The same stream architecture drives the Neo4j graph sync with graceful degradation — DynamoDB mutations are never blocked by graph failures, and agents receive structured fallback hints when the index is unreachable.
By The Numbers
Bottom Line
Enceladus is not a product for sale — it is an operational philosophy made manifest. It demonstrates that a single operator, armed with the right abstractions and governed AI agents, can build, maintain, and learn from production systems at a scale and quality level that would traditionally require a team. The platform's power comes not from complexity, but from the disciplined simplicity of treating every entity, every transition, every mutation, and now every lesson as an opportunity to enforce quality, accumulate wisdom, and align with values at the architectural level.