F

Firefly Transcreate

Context-Aware Multilingual Asset Localization for Adobe Express

Back to Portfolio
Candidate: Souvik Ghosh (IIM Calcutta / B.Tech CSE)
Target Role: PM - Multilingual & GenAI Platform
North Star: Post-Translation Edit Distance (-64%)

Flow Controls Live Simulation

🇮🇳 Hindi (India) High Volume
🇲🇽 Spanish (LATAM) High Volume
🇯🇵 Japanese (Japan) Kanji Density
🇮🇳 Bengali (India) Vernacular
🇩🇪 German (DACH) Text Expansion
Adobe Express Design Canvas Literal Machine Translation
Format: 1200 x 630 (Social Banner)
⚠️ Idiom Error & Bounding Box Overflow
78% (High Edit)
Requires manual overhaul
Overflow (138%)
Text clipped out of bounds
1 / 5 (Poor)
Word-for-word literal mismatch
210 Minutes
Heavy manual review bottleneck

🏗️ Platform Model Routing Architecture

1. Multi-Tiered GenAI Translation Pipeline

Instead of routing all text through generic translation APIs or heavy multi-modal LLMs, Firefly Transcreate utilizes a two-tier model gateway. Short-form headlines use a lightweight domain-adapted model trained on regional commercial copy, keeping response latency under 1.8 seconds.

PM Decision Rationale: High-velocity marketing creators abandon tools if localization takes over 3 seconds. Latency budget caps multi-modal generation to complex visual scenes only.

2. Bounding Box & Typography Auto-Scaler

Languages vary drastically in character density. German text expands by up to 35% compared to English, while Japanese requires larger line-height for legibility. The layout engine dynamically adjusts font size, letter spacing, and line height inside SVG vector text frames upon transcreation payload return.

PM Decision Rationale: Prevents design breakage without requiring manual frame resize by non-designer users, directly protecting exported asset quality.

3. Cultural Context & Festival Knowledge Graph

The prompt engine automatically attaches regional metadata based on target geographic selection (e.g. converting generic "Fall Sale" into "Diwali Dhamaka" in India, or "Golden Week Specials" in Japan).

PM Decision Rationale: Transcreation requires cultural intent matching, not literal dictionary replacement. User research proved 82% of global SMBs struggle with local holiday context.

🛡️ Product Guardrails & Edge Case Matrix

1. Brand Name Protection & Trademark Exclusion

Maintains an immutable Brand Term Lock List. Product names, trademarked slogans, and proper nouns are flagged with zero-transcreation tokens to prevent hallucinated translations of core brand entities.

PM Decision Rationale: Enterprise accounts indicated brand safety as their #1 fear when adopting AI content tools.

2. Low-Confidence Fallback & Manual Override

If the transcreation confidence score drops below 0.82 (e.g., highly specialized legal disclaimers or niche technical jargon), the system highlights the text box in amber and provides a 1-click side-by-side edit panel for human approval.

PM Decision Rationale: Eliminates hallucination risk on sensitive copy while preserving speed for 90% of standard marketing text.

3. Bi-Directional Layout Flipping (RTL Support)

Automatically flips design canvas layouts, alignment anchors, and call-to-action button placements when transcreating into Right-to-Left languages like Arabic and Hebrew.

PM Decision Rationale: True internationalization requires structural layout mirroring, not just text string changes.
Post-Translation Edit Distance Rate
14%
📉 Down from 78% in literal machine translation
Avg Time to Publish (Global Asset)
4.2 min
⚡ Reduced from 210 minutes per campaign
Global Non-English Asset Exports
+42%
🚀 QoQ expansion across international SMBs
Edit Distance & Manual Correction Rate by Market (Lower is Better)
Measured across 5,000 localized creative exports
Hindi (India)
14%
Japanese (Japan)
18%
Spanish (LATAM)
11%
German (DACH)
15%
Legacy Machine API
78%
North Star Metric Tree: Enterprise ARR Expansion

North Star: Enterprise International ARR Expansion ($14.2M Opportunity)

├── Driver 1: Localized Asset Export Velocity (+42%) → Driven by Firefly Transcreate 1-click execution

├── Driver 2: Post-Translation Edit Distance (-64%) → Eliminates manual copy editing friction

└── Driver 3: Global Team Seat Retention (+28%) → Non-English speaking marketing teams adopt Adobe Express natively

Return to Main Portfolio Souvik Ghosh • Product Manager Portfolio