For anyone viewing this: I've continued work on this project and it has come a long way since just the hackathon. My own work uses Cloudflares stack heavily so it was refactored to run my own network, but most of this could be done just as easily continuing on Google Cloud as well.
Major Milestone Update: From GCP Prototype to Production Edge Gateway (PRISM v2) When we originally submitted PRISM (Protocol Routing & Identity Service for MCP) for the All Things Agentic Hackathon, our goal was to solve the Enterprise Zero-Trust & Context Bloat problem: how to deploy autonomous AI agents safely without trusting prompt injection targets, leaking sensitive credentials, or overwhelming model context windows.
Our initial build proved the architecture using Go on Google Cloud Platform (Cloud Run, Cloud Pub/Sub, Firestore, GCS, and Gemini Flash via Vertex AI).
Since the hackathon submission, we’ve taken PRISM from an enterprise concept into a production-grade, daily-driver edge gateway powering real autonomous agents and a 10,000+ note second brain. Here is a rundown of the major architectural upgrades we've shipped:
- Complete Edge-Native Rewrite (Sub-20ms Cold Starts) TypeScript + Hono on Cloudflare Workers: Ported the Go core to a lightweight TypeScript edge worker. Container cold starts on Cloud Run dropped from several seconds to ~18ms edge executions. Cloud-Native Primitive Mapping: Replaced GCP Pub/Sub with Cloudflare Queues (prism-tasks), Firestore with Cloudflare KV (MEMORY_KV), and Google Cloud Storage policies with Cloudflare R2 (POLICIES_R2).
- Native Obsidian Second Brain & R2 Vault Binding Live Vault Operations: Bound production storage directly to Cloudflare R2 (VAULT_R2) via our new R2VaultAdapter. Zero-Friction Journaling & Sync: Agents can now read, write, and quick-capture entries via append_daily_note (which automatically generates the structured 70_Journal/Daily Notes/YYYY-MM-DD.md entry if it doesn't already exist).
- Hybrid Semantic Vector Search at the Edge Edge Embeddings + Supabase pgvector: Built the SupabaseVectorAdapter. Queries compute dense 1024-dimensional embeddings at the edge using Cloudflare Workers AI (@cf/baai/bge-large-en-v1.5) and query Supabase Postgres via RPC (match_doc_embeddings). Sub-Second Knowledge Retrieval: The search_vault MCP tool returns ranked semantic markdown snippets in milliseconds.
- Zero-Trust Input Validation & Self-Healing Loop Edge-Native JSON Schema Enforcement: Added rigorous runtime schema validation (src/validation/schema.ts) for all incoming MCP tool arguments. Directory Traversal & Path Guards: Automatically sanitizes vault paths, enforces relative constraints, and rejects ../ traversal attempts with standardized MCP JSON-RPC -32602 error codes. Meta-Agent Self-Healing: The Gemini reasoning worker now receives full tool JSON schemas in its prompt and features an automated 1-turn self-healing loop—if a tool execution fails due to invalid parameters, PRISM feeds the diagnostic error back to Gemini to auto-correct parameters and retry without crashing the task.
- Gemini 3.8 Flash & Workers AI Multi-Model Bridge Resilient Gemini 3.8 Flash Reasoning: Upgraded the core Meta-Agent (brain_oracle) using the @google/genai SDK with exponential backoff and jitter to gracefully handle transient 429/503 rate limits. Workers AI Bridge (prism-worker-ai): Deployed local edge AI tools behind PRISM via Cloudflare Service Bindings: transcribe_audio via Whisper generate_embeddings via BGE-Large fast_llm_inference via Llama 3.3 70B FP8 Fast vision_ocr via Llama 3.2 11B Vision Instruct OpenAI-Compatible REST Proxy: Exposes edge-native /v1/chat/completions, /v1/embeddings, and /v1/audio/transcriptions endpoints protected by Zero Trust Access.
- Autonomous Identity Mapping (anna_agent) Integrated Cloudflare Access Service Token resolution to map machine identities directly to granular RBAC roles (mobile_agent, desktop_agent, workflow_agent, anna_agent). Continuous autonomous agents run 24/7 with zero-trust least-privilege scoping across tools. Verified by Automated Tests Our edge test suite has grown from a simple sanity check to 16 passing unit and integration tests verifying authentication, RBAC policy isolation, parameter validation, self-healing retries, and multi-model proxying.
Thank you to the Google Cloud & Agentic Hackathon organizers—PRISM started as an ambitious hackathon submission, and it has evolved into the central backbone of our personal agent fleet!
Log in or sign up for Devpost to join the conversation.