Brainforge Knowledge

Eden AI Project Plan

8 min read

Eden AI Project Plan

1. Engagement Overview

1a. Client channels & evaluation

  • Internal (Brainforge): #client-eden-ai — engineering, Linear, staffing.
  • Client-facing (Slack Connect): #external-client-eden-ai — status updates, access requests (GCP, DWD, Slack), demo links, team intros.
  • Stress tests / acceptance: Maintain a shared question corpus (real COO-style prompts) with Danny; run the same set internally before demos. Supports Theme Discovery / weekly pulse acceptance later.
  • Detail: Channel IDs, org-chart field list for identity mapping, and transcript file paths (indexed in eden/transcripts/) live in knowledge/clients/eden/resources/eden-ai-client-channels-and-context.md.

2. Initiatives

Command Center

  • Business unlock: Leverage data from Google Workspace and Slack to uncover themes across the business
  • Delivery date: May 15th
  • Services used: Data / AI / Strategy & Analytics

3. Projects

Data Access + Chat Integration

  • Parent initiative: Command Center
  • Service line: AI / Data
  • Service Lead: @Sam Roberts @Awaish Kumar
  • Start date: March 23rd
  • Target date: April 20th

Theme Discovery and Tailored Reports

  • Parent initiative: Command Center
  • Service line: AI / Strategy + Analytics **
  • Service Lead: @Sam Roberts + @Robert Tseng
  • Start date: April 20th
  • Target date: May 15th

4. Milestones

M1: Danny’s Slack + Danny’s Google workspace can be interacted with via chat

  • Parent project: Data Access + Chat Integration
  • What the client receives: as titled
  • Target date: April 6th

M2: Eden’s Google Workspace can be interacted with via chat

  • Parent project: Data Access + Chat Integration
  • What the client receives: as titled
  • Target date: April 13th

M3: Eden’s Google Workspace + Slack workspace can be interacted with via chat w/ Anonymization

  • Parent project: Data Access + Chat Integration
  • What the client receives: as titled
  • Target date: April 20th

M4: Theme Discovery Baseline Delivered

  • Parent project: Theme Discovery and Tailored Reports
  • What the client receives: baseline theme output package showing first-pass discovered themes, sample redacted signals per theme, and a short interpretation of what each theme appears to represent.
  • Target date: May 4th

M5: Theme Quality Validation Completed

  • Parent project: Theme Discovery and Tailored Reports
  • What the client receives: Themes that are reviewed and tuned with Eden so labels are business-readable, overlap is reduced, and noise is reduced.
  • Target date: May 7th

M6: Bandwidth and Friction Signals Report Delivered

  • Parent project: Theme Discovery and Tailored Reports
  • What the client receives: a report that shows clear theme-level signals, including bandwidth load, talk-to-task balance, and friction indicators, with interpretation guidance that is easy to understand and act on.
  • Target date: May 11th

M7: Weekly Pulse Report v1 Delivered

  • Parent project: Theme Discovery and Tailored Reports
  • What the client receives: An actual email generated from production data. In addition to the prior report format, the output includes clear week-over-week and month-over-month trends.
  • Target date: May 14th

M8: Command Center v1 and Handoff Delivered

  • Parent project: Theme Discovery and Tailored Reports
  • What the client receives: The finalized report delivered to their inbox, a repeatable workflow, and all documentation needed to operate and extend the process is presented.
  • Target date: May 15th

5. Technical Approach

Spike Notes

Data Access + Chat Integration

Relevant Playbook/s:

Target Process (Step-by-Step)

Purpose: Build privacy-first across Eden’s four workspace systems, mapping raw user identities to stable role-based tokens before any data lands in LLM context.

  1. Source authentication — Provision a GCP service account with Domain-Wide Delegation and create a Slack app with search/history scopes. Eden IT approves access.
  2. Identity anonymization layer — Build a secure mapping table that resolves all user identities to stable role-based tokens (e.g. Provider_A, Ops_Tech_1) and a PII redaction middleware that strips real names/emails before any data reaches the LLM or the COO.
  3. Slack data access tools — Build agent tool functions for Slack search, thread reading, and channel stats using the RTS API / MCP. All results anonymized before synthesis.
  4. Google Workspace data access tools — Build agent tool functions for Drive search, Drive Activity, file comments, Gmail metadata, Calendar events, and user directory using GWS CLI with DWD. All results anonymized.
  5. Cross-platform orchestration — Build the orchestration agent that plans which tools to call for a given question, executes Slack + GWS queries in parallel, and synthesizes a unified anonymized answer.
  6. Custom UI: scaffold and chat — Build the Next.js web app with Google OAuth login, streaming chat interface, and an API layer that proxies to the agent and serves dashboard data.
  7. Deploy and validate — Deploy agent to Vertex AI and web app to Vercel/Railway. Validate with Danny across single-source, cross-platform, anonymization, and dashboard test scenarios.

Tools & Stack

  • Primary tools: gws cli, slack rts api/ mcp, mastra/adk, nextjs
  • Integrations: Google Calendar API, Gmail API, Google Drive API, Slack API,
  • Infrastructure: GCP for agent and UI

Architecture Notes / Dependencies

  • Anonymization must be applied at the extraction layer — raw PII cannot touch the LLM context at any stage;
  • Demos vs org-wide access: Early milestones may credibly use end-user OAuth (e.g. Danny’s workspace login) in the chat UI while org-wide answers still rely on service account + Domain-Wide Delegation + impersonation. Milestone language should not be read as “user scope only” if the product goal remains cross-org visibility — see risks below.

Performance / Accuracy Targets

  • N/A if not applicable

Resourcing

  • 1 AI engineer 15-20 hours/week

Risks

  • Google Workspace domain-wide delegation approval delayed → Calendar + Gmail access blocked at kickoff
  • User credentials vs service account: Progress on M1–M2 can look “done” under a single user’s OAuth while M3 (cross-workspace + anonymization) still depends on DWD/SA path — track both explicitly so client demos do not confuse the two modes.

Theme Discovery and Tailored Reports

Relevant Playbook/s:

  • Unsure

Target Process (Step-by-Step)

Purpose: Leverage the data access layer built in Project 1 to automatically discover recurring themes across Eden’s Slack and GWS, then generate a scheduled report that tracks those themes and how they shift week-over-week.

  1. Theme Extraction Baseline — Run the Project 1 agent against the now-live Slack + GWS data to pull a first-pass corpus of signals. Apply Robert’s theme analysis framework as the clustering/labeling template. Output: discovered theme list, sample redacted signals per theme, plain-language interpretation of each theme. Store outputs in a structured format (TBD: Supabase or GCS) as the baseline for future comparison. (→ M4)
  2. Theme Validation with Eden — Working session with Danny to review and tune the baseline themes: make labels business-readable, reduce theme overlap, remove noise. Update theme taxonomy in the structured store. (→ M5)
  3. Bandwidth and Friction Signals Report — Build a report layer over validated themes. Compute: bandwidth load per theme, talk-to-task balance, friction indicators. Deliver as a structured report with interpretation guidance. (→ M6)
  4. Scheduled Report Agent — Build or extend the Mastra agent to run on a weekly cadence. On each run: query Slack + GWS via the same tools from Project 1 → extract signals → classify into themes → generate report → store snapshot → use prior snapshot as context to compute WoW/MoM trends. Two-agent pattern: Agent A generates the current report, Agent B compares to the prior period. Deliver as email from production data. (→ M7)
  5. Handoff and Documentation — Finalize the repeatable workflow, document how to operate and extend, deliver final report to Danny’s inbox. (→ M8)

Tools & Stack

  • Primary tools: Mastra agent, GWS CLI, Slack MCP / RTS API (same as Project 1), scheduled trigger (Cloud Run job or Dagster schedule)
  • Integrations: Google Calendar API, Gmail API, Google Drive API, Slack API
  • Infrastructure: GCP (agent + scheduler), report delivery via Gmail API
  • Storage: TBD — Supabase or GCS for theme outputs and report snapshots (decision needed before M7)

Architecture Notes / Dependencies

  • Project 1 (M3) must be live before meaningful theme extraction can begin — M4 is intentionally gated on real data access
  • Storage decision for theme snapshots gates the comparison/trending feature — must be resolved by end of M5
  • CLI/MCP route is the first option for report generation; only build a light ingestion pipeline if historical data depth via CLI is insufficient
  • Theme quality depends on Eden participation in the M5 validation session

Performance / Accuracy Targets

  • N/A — to be defined after M4 baseline is reviewed with Eden

Resourcing

  • 1 AI engineer — agent, scheduling, storage, delivery pipeline
  • 1 Analyst — theme framework template and theme quality validation (M5)

Risks

  • Historical data depth in GWS Drive/Docs is unknown until DWD is live — if insufficient, will work with Awaish + Robert to ingest the right data
  • CLI/MCP route may not support scheduled/historical report generation at scale → fallback is a light ingestion pipeline
  • Hard to fully spec until Project 1 data access is live — this technical approach is intentionally high-level and will be refined after M3

6. Open Questions

CSO Questions (Client / Scope)

SL Questions (Technical Blockers / Unknowns)

  • How deep does GWS historical data go via the CLI + DWD? Specifically Drive/Docs — if insufficient, do we need to work with Awaish + Robert to build a light ingestion pipeline for more data?
    • Resolved, historical snapshots of themes that are saved will suffice

Head of Delivery

7. Sign-offs

  • CSO - Business scope, milestones, and client accountability
    • Signed off?
  • SL - Technical approach, effort estimates, and delivery accuracy
    • Signed off?
  • Head of Delivery - Feasibility, business case, and leadership alignment
    • Signed off?
  • Client sponsor name - Client alignment and acceptance of scope + timeline
    • Signed off?

Reference Documents

  • Implementation plan template: vault/gtm/services/templates/implementation-plan-template.md
  • Playbook §8.2 — Atomic chain: knowledge/standards/PLAYBOOK_DEVELOPMENT_PLAN.md
  • Client channels, eval corpus, org fields, transcript pointers: knowledge/clients/eden/resources/eden-ai-client-channels-and-context.md

Graph View