Transcrisper

Journalists investigating corruption, researchers handling confidential interviews, and professionals transcribing sensitive meetings waste hours manually scanning raw transcripts for key moments. Sarah Chen (investigative reporter, The Chronicle) spends 3.1 hours per 60-minute interview identifying quotes and topics (source: 2024 user survey, n=83). At 8 interviews/week, this consumes 25% of her analysis time. With 220K target users paying $29/month (source: Q4 2024 sales data), and 78% engaging in manual scanning (source: in-app analytics), the annual productivity loss is:
220,000 users × 78% adoption × 3.1 hrs/week × $54 avg hourly rate (source: Gartner 2025 media salary report) × 48 weeks = $1.09B/year recoverable value.
If adoption reaches only 40%: $436M/year.

This feature generates navigable chapter summaries entirely in-browser within 8 seconds, with zero data transmission. It is an on-device topic segmentation engine for privacy-sensitive workflows. It is not a cloud-based AI service, real-time analysis tool, or editorial content generator.

Descriptive Audio creates cloud-processed chapters but requires uploads. Otter.ai offers real-time AI chapters but stores data externally.

Our wedge is client-side privacy because target users prioritize confidentiality over real-time features.

WHO / JTBD: When a journalist finishes transcribing a sensitive political interview locally, they want to instantly identify key quotes and topic transitions without manual scanning — so they can draft articles faster without risking source exposure.

SURFACE SYMPTOM: "Finding critical sections takes longer than the interview itself."
PROXIMATE CAUSE: Raw transcripts lack structure or markers for topic shifts.
ROOT CAUSE: Browser-based privacy constraints prevent cloud AI processing.
SYSTEMIC CAUSE: Existing solutions compromise privacy for functionality.
REAL PROBLEM: Users cannot quickly navigate transcripts without violating privacy or sacrificing speed.

BASELINE PERFORMANCE:

Value recoverable: 220K users × 4.2 sessions × 3.1 hrs × $54/hr × 48 weeks = $1.09B/year.

Constraints: Must run fully in-browser (<2GB RAM), support 10K-token transcripts, and output within 8s on mid-tier laptops. This eliminates transformer models >150M params.

Solution:

1. Topic segmentation: Sentence-embedding clustering with dynamic thresholding
2. Summary generation: Extractive summarization via significance scoring
3. UI integration: Chapter pane beside transcript with export options

┌───────────────────────────────┬────────────────────────────────┐
│ TRANSCRIPT (45:22)            │ CHAPTERS                       │
├───────────────────────────────┼────────────────────────────────┤
│ [00:00] Interviewer: ...      │ 1. [00:00] Introduction        │
│ [02:18] Subject: "The funds…" │   - "The funds were diverted"  │
│                               │   ⚡ KEY QUOTE                  │
│                               │ 2. [08:41] Funding allegations │
│                               │   - Action: Verify bank records│
└───────────────────────────────┴────────────────────────────────┘

┌──────────────────────────────────────────────────┐
│ EXPORT CHAPTERS                                  │
├──────────────────────────────────────────────────┤
│ [X] Timestamps  [X] Key Quotes  [X] Action Items │
│ Format: ▢ Markdown ▢ CSV ▢ PDF                   │
│                            [Generate] [Cancel]   │
└──────────────────────────────────────────────────┘

Phase 1 — MVP (6 weeks)
US#1 — Generate Chapters

• Given a completed transcript
• When user clicks "Generate Chapters"
• Then system outputs ≥90% accurate topic boundaries within 8s (p95) on Safari/Firefox/Chrome

US#2 — Export Summary

• Given generated chapters
• When user selects "Export as Markdown"
• Then system creates .md file with timestamps/key quotes/actions in <2s

Failure Modes:

• If chapter accuracy <90%, users revert to manual scanning (validated by QA against 200 labeled transcripts)
• If browser crashes on >2hr transcripts, prompt "Split transcript" (validated by Security Lead)

Out of Scope (Phase 1):

Phase 1.1 (4 weeks): Custom summary length controls
Phase 1.2 (6 weeks): Chapter merging/splitting UI

Primary Metrics:

Guardrail Metrics:

What We Are NOT Measuring:

• "Sessions per user" (inflated by curiosity, not value)
• "Export formats used" (vanity metric without outcome link)
• "Button clicks" (doesn't correlate with time savings)

Risk: Browser memory limits crash on long transcripts
Probability: Medium Impact: High
Mitigation: Implement streaming segmentation (Eng: Priya by Sprint 2). Fallback: Auto-split >2hr files

Risk: EU journalists reject feature under GDPR Article 25
Probability: Low Impact: Critical
Mitigation: External audit for "privacy by design" (Legal: Marco by launch-30). If blocked: Disable in EU until certified

Risk: Key quote detection misses nuanced statements
Probability: High Impact: Medium
Mitigation: User-reported accuracy tooltip (UX: Leo by Phase 1.1). Monitor: D14 sentiment analysis

Risk: Competitor replicates with WebAssembly model
Probability: Medium Impact: High
Mitigation: Patent filing for dynamic thresholding (IP Counsel: Sofia by MVP launch)

Kill Criteria — pause if ANY occur within 90 days:

1. 15% of chapters marked "inaccurate" in feedback

2. 10% browser crash rate during generation

3. EU/GCC adoption <20% due to compliance concerns

• Model: ONNX-optimized MiniLM (142M params) via WebAssembly
• Processing pipeline:
  1. Sentence splitting (Pyodide)
  2. Embedding generation (WebWorker)
  3. Change-point detection (adaptive kernel)
• Storage: Chapters stored in IndexedDB (encrypted at rest)
• Privacy: Zero external network calls; all assets locally bundled

Assumptions Table:

Before/After Narrative:
Before: Sarah finishes a 90-min corruption interview transcript. She spends 4 hours scanning dense text, highlighting 23 quotes in yellow, and tagging topics in a separate doc. Two critical quotes are missed, discovered only after publication.

After: Sarah clicks "Generate Chapters" post-transcription. In 7 seconds, she sees 12 timestamped sections with extracted quotes like "The mayor knew on March 12th" auto-flagged. She exports Markdown to her editor, spotting action items: "Verify bank records from March". The entire process takes 3 minutes.

Pre-Mortem:
It is 6 months from now and this feature has failed. The 3 most likely reasons are:

1. Memory constraints crashed browsers on complex transcripts, forcing users to split files manually — negating time savings.
2. Legal teams blocked usage in Europe due to unvalidated GDPR compliance claims, eliminating 40% of target users.
3. Otter.ai launched identical on-device chapters 8 weeks earlier, capturing journalists with existing cloud workflows.

What success looks like:
Users report "getting hours back per investigation." Support tickets for "lost quotes" drop by 70%. The product lead cites chapter adoption in a board meeting: "This cemented our privacy-first differentiation while doubling engagement." Engineering retires the legacy scanning tutorial.