TITAN/JARVIS — Chief of Staff Roadmap (post-2026-05-14)

> Scout-researched, founder-grade backlog of newsletter + dashboard enhancements.

> Status: shipped today vs queued.

---

✅ Shipped today (visible now)

Memo browser — 541 memos · Sonnet-rendered HTML view · .docx download

https://innerverse-voice-scratch.s3.us-east-1.amazonaws.com/research/

Email open-rate leaderboard — top + bottom performers, by-day spark, 7/14/30/90d window picker

https://innerverse-voice-scratch.s3.us-east-1.amazonaws.com/research/stats/index.html

Follow-up debt tracker — Sonnet-extracted overdue commitments from PENDING + memos. Today: 16 overdue items surfaced including Q33 SnapDomain repo path and 4 staging-queue runs you said you'd do

https://innerverse-voice-scratch.s3.us-east-1.amazonaws.com/research/followup/index.html

AWS cost tracker — weekly spend with WoW delta + month-end projection. Currently $18.83/wk, -5% vs prior week

https://innerverse-voice-scratch.s3.us-east-1.amazonaws.com/research/cost/index.html

On-demand narration narrate.py + Lambda + dialogue button on every newsletter player

YouTube oEmbed enrichment (pre-pass in research_indexer) + Playwright screenshot capture (enabled in all 5 newsletter crons)

Daily crons registered for both new dashboards (titan-followup-debt-daily at 10:00 UTC, titan-cost-tracker-daily at 10:30 UTC)

---

🚧 Queued — Newsletter enhancements (ranked by impact/effort)

1. Audio chapters with timestamps (S · 1-2 hours)

What: Pre-defined chapter structure per narration (00:00 TL;DR, 01:30 Big story, 04:00 Quick hits, etc.) emitted as VTT alongside the MP3. Player + email gets clickable timestamp links.

How: Modify narrate.py to:

Have Sonnet output chapter offsets when generating the dialogue/solo script
Track byte-count → time conversion per Polly chunk (Polly returns audio at known kbps)
Write .vtt alongside .mp3 on S3
Player page (build_player_page.py) renders chapter list with audio.currentTime = X onclick

Why: 2-3× completion rate per Scout's findings. Busy listeners self-route.

---

2. Personalized "For You" section per recipient (M · 1 day)

What: Each subscriber gets 2-3 dynamically chosen "For You" items at the top of the newsletter based on their click history.

How:

New DDB table titan-reader-interests keyed on email_hash with vector of clicked topics (last 30 days)
SES click tracking (already half-built via open pixel — extend to wrap CTAs with redirect Lambda)
At send time, for each recipient: Bedrock Sonnet ranks today's stories against their interest vector, picks top 3
Renders into a <personalized> block in the email template

Why: Most newsletters with 10× the team haven't shipped this. Differentiator.

Cost: ~$0.005/recipient × N subscribers × 30 = roughly $5-15/mo for typical TITAN scale.

---

3. Reader poll → next-issue editorial signal (S · 2 hours)

What: One question at the bottom of every newsletter. Results published in next issue: "You voted 67% X, so this week we cover Y."

How:

Embed Tally.so or self-hosted form pointing to a new /poll-vote Lambda
DDB table titan-poll-votes keyed by (issue_slug, date, option)
New script build_poll_recap.py pulls yesterday's results, injects into today's template header

Why: Drives open-rate spike on follow-up issue; builds habit loop.

---

4. Interactive AI Q&A grounded in issue archive (M · 1-2 days)

What: A chat endpoint on each newsletter player ("Ask anything about this week's coverage") grounded only in past issues via RAG.

How:

Bedrock Knowledge Base over s3://innerverse-voice-scratch/newsletter-archive/
New Lambda /ask-archive → Sonnet with KB attachment
Add chat widget to player page (reuse JARVIS voice UI components)

Why: Transforms static archive into a tool. Stops competing with ChatGPT, starts being the expert layer on top of it.

---

5. Telegram flash alerts for breaking AI news (S · 3-4 hours)

What: Separate Telegram channel that fires within 1 hour of confidence-scored breaking news. Super-fan retention play.

How:

Telegram Bot API key (free) → titan-telegram-poster Lambda
ORACLE feed ingest already runs hourly — add hook: if confidence >= 0.85 AND story is freshness < 2h, post to Telegram channel
Channel link in newsletter footer as "join for breaking drops"

Why: Solves the "I unsubscribed because the daily wasn't fast enough" failure mode without breaking the daily cadence.

---

6. Linked data drops (CSVs, datasets, benchmarks) (S · per-issue)

What: When an issue covers a benchmark/dataset, attach the CSV (or Google Sheet link) with a 5-line usage guide.

How: Extract dataset URLs during research-index step → fetch first ~1000 rows → store at /datasets/<slug>/<date>.csv → inject link into email template.

Why: Practitioners bookmark the newsletter as a tool, not just content.

---

🚧 Queued — JARVIS Chief-of-Staff widgets (ranked)

7. Yesterday's metrics vs trailing-7d average (S · 2 hours)

What: Daily widget on the tasks dashboard showing key metrics (newsletter opens, signups, audio plays, AWS spend) vs 7-day baseline.

How: Single new script daily_metrics_snapshot.py pulls from existing sources (SES, DDB, Polly, Cost Explorer), writes a JSON snapshot, renders a strip on the main dashboard.

Why: Removes the "how are things going?" decision tax every morning.

---

8. Ideas raised in chat → orphaned task detector (M · 1 day)

What: Daily diff: scan recent JARVIS conversation logs for action language ("I should...", "let's ship X", "add Y to...") that has no corresponding PENDING.md entry. Surfaces "orphaned ideas."

How:

Pull last 24h of titan-chat-sessions DDB rows
Bedrock classifier flags action-intent messages
Diff against PENDING.md headlines
Unmatched → daily section in followup-debt page

Why: The most expensive lost work is the thing you said out loud but never wrote down.

---

9. Time allocation vs strategic intent (M · 1 day)

What: Set target split once (e.g., 50% build, 30% distribution, 20% ops). JARVIS infers category weekly from task tags + commits/SES sends and reports actual vs target.

How: New script weekly_allocation_report.py ingests git commits, task tags, SES sends → Bedrock classifies → renders a single bar chart + 1-line commentary.

Why: Solo founders drift into reactive mode invisibly.

---

10. Staging memory queue monitor (S · 1 hour)

What: Banner on dashboard showing how many ORACLE-generated insights are in staging unreviewed + age of oldest.

How: wc -l F:/TITAN/knowledge/staging/* + oldest mtime → JSON to S3 → widget pulls.

Why: Memory system value degrades silently if staging fills.

---

11. Nightly report delta-only view (S · 2 hours)

What: Widget showing only what's changed between last night's report and the prior night — not full reread.

How: Diff two nightly reports → Sonnet summarizes deltas in 3 bullets.

Why: Turns a 5-min read into a 30-sec scan.

---

🎯 Recommended next session order (impact/effort)

1. Audio chapters — visible UX win, Polly + narrate.py already running

2. Reader poll — easiest engagement loop, no new infra

3. Telegram breaking-news bot — super-fan retention, single Lambda

4. Personalized For You section — biggest brand differentiator

5. Interactive AI Q&A — flagship "the archive is a tool" feature

6. Orphaned task detector — directly compounds the follow-up debt scanner

7. Time allocation report — for when the build vs distribution drift matters

---

Notes

All four of today's shipped dashboards are reachable via /research/index.html button row
Cron will keep follow-up debt + cost dashboards fresh daily without intervention
Memo browser cards regenerate when you re-run build_research_memos_page.py (no auto-cron yet — add if memo count keeps growing)
Open question: route /research/ through Cloudflare worker once you provide a CF API token (currently served via direct S3 URL — works, just longer URLs)