EMMANUEL_ARO_/>← Back to work

// CASE STUDY

Dark Phoenix — AI Podcast Clipper (LunarTech)

Delivered LunarTech's AI podcast video clipper end-to-end: engineered the full GPU processing pipeline on Modal (TalkNet active-speaker detection → WhisperX transcription → Gemini AI moment selection → ffmpeg clip render), YouTube ingestion via yt-dlp, provider-agnostic S3 object storage (Backblaze B2), and a branded watermark burned directly into exported MP4s. Resolved 8 critical production deployment failures across Modal, WhisperX, Vercel, and Next.js to bring the platform from broken to fully shipped.

Role
Full-Stack AI Engineer
Year
2026
Context
LunarTech
Status
Shipped & Active
PythonModal (GPU)WhisperXTalkNetGemini AINext.jsSupabaseffmpegAWS S3
01.

Overview

Dark Phoenix is an end-to-end AI podcast clipping engine that ingests long-form video, analyzes it for viral moments using generative AI, and renders high-quality short-form clips (with subtitles and watermarks) entirely in the cloud.

02.

The problem

Producing short-form clips from long podcasts is a highly manual, labor-intensive process requiring video editors to watch hours of content. The previous iteration of the platform was fundamentally broken, suffering from massive GPU timeouts and deployment failures.

03.

My approach

I took ownership of the entire stack, systematically resolving 8 critical production failures across Modal, WhisperX, Vercel, and Next.js. I rebuilt the ingestion pipeline using yt-dlp and routed all heavy AI processing through a highly optimized, asynchronous GPU pipeline.

04.

Architecture decisions

Asynchronous GPU orchestration — Deployed the intensive AI workloads (TalkNet active-speaker detection and WhisperX transcription) on Modal to ensure scalable, on-demand GPU provisioning without locking up the main server.

AI-driven moment selection — Leveraged Gemini AI to contextually analyze the highly-accurate transcripts and isolate the most engaging segments.

Headless rendering — Engineered a server-side ffmpeg pipeline to dynamically crop vertical video, burn in branded watermarks, and upload the final MP4s to Backblaze B2.

05.

Outcome

The platform was successfully rescued from a broken state and launched into production. It now autonomously processes hours of podcast footage into highly engaging, ready-to-publish social media clips in a fraction of the time.

View live ← All projects