AITrailblazer AI Agent Publishing

Turn any expert publication archive into MongoDB-backed agent memory with TripCodes, Rivers, and Gemini-powered research workflows.

Judge Quickstart

1. Open the hosted demo.
2. Enter Trip Code: HUT-RIVER-001.
3. Click Run Judge Demo.
4. Confirm Gemini / Vertex AI synthesis, Agent Builder retrieval lane, MongoDB memory objects, MongoDB MCP tool-call proof, and reader-session follow-up memory.

Smoke tests: GET /health and GET /v1/judge-demo.

Problem

Expert publishers have years of valuable articles, but readers and agents cannot reliably reuse them. Static archive search does not preserve claims, updates, evidence trails, or follow-up context.

Solution

AITrailblazer turns expert publication archives into agent memory. A reader enters a Trip Code from an article, MongoDB resolves the article, claims, River links, and session memory through MCP, and Gemini synthesizes a grounded answer that can continue across follow-up questions.

DeltaSignal is the proof archive; the product is the generalized publisher-memory agent.

Demo Flow

1. Resolve HUT-RIVER-001 through MongoDB-style TripCode, claim, River, and three article records.
2. Show the agent plan, source context, River memory, Gemini answer, next actions, and runtime proof.
3. Ask a follow-up and show second-turn memory from the reader session.

Hosted Runtime

The judge runtime is deployed on Cloud Run at https://aitrailblazer-ai-agent-publishing-rmycwek6ba-uc.a.run.app. The hosted service exposes GET /health and GET /v1/judge-demo for a fresh-browser or curl smoke test. Current hosted proof invokes Gemini live through Vertex AI and invokes the Agent Builder lane through a Google Cloud Agent Search engine over the three seeded River documents.

Public Repository

The competition-facing public source repository is https://github.com/aitrailblazer/aitrailblazer-ai-agent-publishing. It contains only this sanitized publishing workspace, not the private parent monorepo.

Usage

AITrailblazer is a publication-memory agent. It lets a reader enter a known TripCode from an article, newsletter, or research note and retrieve the article's source context, claim cluster, River memory, and follow-up path.

Example TripCode Flow

1. Open the hosted Cloud Run app.
2. Enter a known TripCode: HUT-RIVER-001.
3. The agent resolves the TripCode through MongoDB.
4. MongoDB MCP retrieves three linked River articles, claims, River edges, and prior memory.
5. Gemini generates a grounded synthesis from the retrieved publisher memory.
6. The user asks a follow-up: What should I monitor next?.
7. The agent uses the saved reader session to continue without requiring the TripCode again.

Demo Questions

• Resolve HUT-RIVER-001
• What changed across this River?
• What should I monitor next?
• Show the claims behind this article.
• Which connected article updates this claim?

What This Proves

AITrailblazer proves that a publication archive can become agent memory. MongoDB stores the memory objects, MCP exposes them to the agent, Gemini synthesizes the answer, and Google Cloud hosts the working application.

Public Safety Boundary

• No credentials.
• No Devpost cookies.
• No Google Cloud keys.
• No billing exports.
• No private screenshots.
• Atlas URI only through Secret Manager.
• Demo seed data only.
• Delta Signal used as proof archive, not a private data dump.
• No claims that this submission is a live SEC/XBRL market-data product.
• No investment recommendations, paid-route pricing claims, or trading signals.

Scope Boundary

This Devpost project is AITrailblazer AI Agent Publishing: a publisher-memory agent that proves Trip Code resolution, River memory, MongoDB MCP retrieval, Gemini synthesis, and second-turn session memory. It does not publish or depend on any separate Delta Signal product surface, live SEC/XBRL readiness contract, paid access route, or investment-signal workflow.

Target Users

• Independent publishers.
• Research firms.
• Analyst teams.
• Technical blogs.
• Newsletters.
• Education archives.
• Financial research teams.

Why It Matters

AITrailblazer lets old content become usable memory for future agents. Instead of a static archive, the publisher gets a living claim graph with continuity, updates, contradictions, and follow-up sessions.

Architecture

Cloud Run HTTP service, Secret Manager-ready configuration, MongoDB Atlas data model, MongoDB MCP context surface, Gemini synthesis lane, and deterministic fallback for local tests.

Entrypoint

index.html is the GitHub Pages-compatible publishing surface. It is a self-contained StrategiX Visual Spec with an embedded XML contract.

Start Here And Video Slides

START_HERE.html is the judge command center modeled after the DeltaSignal demo style: open the hosted root, verify health, run judge proof, resolve HUT-RIVER-001, and test second-turn memory. VIDEO_SLIDES.html is the companion under-three-minute recording storyboard.

Public Docs

Public docs is the single consolidated StrategiX Visual Spec for the submission. It covers the product contract, competition requirements, MongoDB and MCP architecture, build checklist, Atlas mode, sanitization posture, and the remaining video-demo submission gate.

Devpost Submission Pack

DEVPOST_SUBMISSION.html contains copy-ready Devpost fields: title, tagline, description, technical implementation, impact, judging notes, built-with tags, submission links, and final checklist.

Final Video ExecPlan

EXECPLAN_FINAL_SUBMISSION_VIDEO.html contains the public sanitized implementation plan and recording runbook for the final under-three-minute Devpost demo video. Use make video-url-check DEMO_VIDEO_URL=https://... after the public YouTube or Vimeo link exists.

Required Technologies Used

• Gemini synthesis lane for grounded answers.
• Google Cloud Agent Builder or Gemini Enterprise Agent Platform as the orchestration lane.
• MongoDB data model for articles, TripCodes, claims, River edges, reader sessions, and agent runs.
• MongoDB MCP Server context path for database-grounded agent access.
• Cloud Run deployment target.

MongoDB Track Implementation

The required collection model is articles, tripcodes, claims, river_edges, reader_sessions, and agent_runs. The runtime exposes this model in GET /v1/judge-demo and calls the official MongoDB MCP find tool against a seeded MongoDB store. Set MONGODB_DEPLOYMENT_KIND=atlas and override MDB_MCP_CONNECTION_STRING for an external Atlas database.

Agent Builder and Gemini Implementation

The judge-proof path is: input TripCode, plan retrieval steps, use MCP context, synthesize with Gemini, write reader-session memory, then answer a follow-up using prior River state.

MCP Server Usage

The MCP proof should show schema, collection, retrieval, or tool-call evidence. The runtime response includes a Runtime Proof panel so judges do not have to infer partner usage.

Use the official MongoDB MCP server through mongodb-mcp-server@latest. Configure external Atlas credentials only through environment variables or Secret Manager, then run make atlas-check before make deploy-atlas. The Cloud Run image starts MongoDB in embedded mode, or skips local MongoDB in Atlas mode, seeds the judge demo collections when enabled, launches official MCP HTTP transport in read-only JSON mode with externally managed sessions, and points MCP_SERVER_URL at http://127.0.0.1:3000/mcp. The app sends Accept: application/json, text/event-stream, which is required by the official HTTP transport.

Local Commands

make check
make run
make demo
make validate
make preview
make mcp-check
make atlas-check
make video-url-check DEMO_VIDEO_URL=https://...
make submission-readiness
make mcp-run
make deploy-atlas ATLAS_SECRET=publishing-mongodb-atlas-uri

Run make submission-readiness before any Devpost final submit. It executes the local gates, validates the official MongoDB MCP prerequisite, optionally runs Atlas preflight when Atlas mode is configured, scans public files for prohibited private evidence/token patterns, checks the hosted root, health endpoint, consolidated docs, license, judge proof, TripCode resolve flow, second-turn follow-up memory, public GitHub repo, raw license, and raw consolidated docs.

Runtime Shape

The Go service mirrors the useful shape of ../deltasignal-ai-agent: a small Cloud Run HTTP surface, deterministic demo mode, judge-friendly curl script, Dockerfile, Cloud Build template, and explicit public/private boundaries.

Local endpoints include GET /health, POST /v1/archive-brief, POST /v1/tripcode, GET/POST /resolve, GET /v1/judge-demo, and GET /v1/usage.

Deployment Setup

Use make deploy-source or make deploy after setting PROJECT, REGION, and the required environment variables. Use a separate Google Cloud project for this competition and link it to the billing account with the available grant credits.

For Atlas-backed proof, create a Secret Manager secret containing the Atlas connection string, then run make deploy-atlas ATLAS_SECRET=publishing-mongodb-atlas-uri. The deploy target maps that secret to MDB_MCP_CONNECTION_STRING and sets MONGODB_DEPLOYMENT_KIND=atlas without writing the URI to source.

Environment Variables

• PORT: local or Cloud Run port.
• PUBLISHING_DEMO_API_KEY: optional demo protection key.
• PUBLISHING_USE_GEMINI: production switch for Gemini runtime path.
• GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_LOCATION, GOOGLE_GENAI_USE_VERTEXAI: Google Cloud runtime settings.
• MONGODB_URI, MONGODB_DATABASE, MONGODB_DEPLOYMENT_KIND, MONGODB_SEED_DEMO_DATA: Atlas or embedded MongoDB settings.
• MDB_MCP_CONNECTION_STRING: official MongoDB MCP connection string input.
• MDB_MCP_API_CLIENT_ID, MDB_MCP_API_CLIENT_SECRET: optional Atlas service-account path for official MongoDB MCP tools.
• MCP_SERVER_URL, MCP_METHOD, MCP_TOOL_NAME, MCP_SESSION_ID: app runtime path to official MongoDB MCP HTTP transport.
• MDB_MCP_READ_ONLY, MDB_MCP_HTTP_RESPONSE_TYPE, MDB_MCP_EXTERNALLY_MANAGED_SESSIONS: MCP server safety and HTTP compatibility settings.

Data Model

The judge dataset should use three seeded River articles: a prior thesis article, the anchor TripCode article, and a follow-up monitor-next article. Keep the dataset to 1 TripCode path, 3 to 9 claim nodes, 2 to 4 River edges, and 1 to 2 sessions. Do not ingest the full DeltaSignal archive for the first submission.

Screenshots

Final screenshots should show the TripCode input, source context, River memory, Gemini answer, next actions, and runtime proof panel.

Demo Video

Keep the video under three minutes and make proof visible in this order: problem, solution, live Trip Code demo, runtime proof, follow-up memory, impact, and close.

1. 0:00-0:20: expert publishers have years of articles, but readers cannot reuse claims, updates, or follow-up context through normal archive search.
2. 0:20-0:45: AITrailblazer turns a publication archive into MongoDB-backed agent memory using Trip Codes, River links, claims, sessions, MongoDB MCP, Gemini, and Cloud Run.
3. 0:45-1:25: enter HUT-RIVER-001 and show the resolved article, three River articles, claim cluster, River edges, and source context.
4. 1:25-1:55: show Gemini live, Agent Builder live, MongoDB live, MongoDB MCP live, and Cloud Run hosted proof.
5. 1:55-2:25: ask What should I monitor next? and show the agent continuing from saved reader-session memory.
6. 2:25-2:50: explain how this generalizes to newsletters, research firms, technical blogs, local journalism, financial research, and any publisher with a valuable archive.
7. 2:50-3:00: close with MongoDB as the memory layer, MCP as the agent access path, Gemini as the synthesis layer, and Cloud Run as the live runtime.

License

The repository includes an OSI-approved MIT LICENSE at the root.

Compliance Notes

Contest-facing code, docs, screenshots, configs, commits, and video should contain only permitted Google Cloud and MongoDB evidence. Do not publish private planning files or local credentials.

The public submission should stay bounded to the seeded publishing-memory demo. Do not add claims about live market-data coverage, SEC/XBRL production readiness, paid access pricing, or investment recommendations unless separately verified and intentionally included in a future product surface.

Source Inspiration

The structure is inspired by ../deltasignal-ai-agent: static public page first, explicit competition positioning, and clear separation between public presentation and private runtime credentials. The new submission should keep DeltaSignal as the proof case while pitching a broader publisher product.

MongoDB Fit

The Rapid Agent MongoDB resource page describes MongoDB Atlas as a persistent memory layer for AI and agentic workloads, combining operational, vector, and semantic data. That makes MongoDB the closest partner lane for publication archives, TripCode objects, River continuity, vector search, and MCP-backed database access.

Boundaries

This repository should not contain Devpost cookies, Google Cloud keys, billing exports, backend secrets, or private screenshots. Public claims should link back to the Rapid Agent resources page, the submission runtime, or checked local evidence.