supernova

Inspiration

Every day creators are under pressure to push out fresh, high–quality short-form video for TikTok, Reels, Shorts, and LinkedIn. We kept hearing the same pain points: “Good ideation already takes forever.” “Recording and captioning burns hours.” “I don’t have an editor on call.” We envisioned a “one-click content reactor” that: reads the market in real time, tailors the script to the creator’s own voice, records an AI avatar (so no camera pressure/anxiety), autogenerates dynamic captions + B-roll, exports a fully-edited 9:16 video ready for upload. That is supernova; a vertical-video assembly line driven by LLMs, LangChain agents, social-graph scraping, the HeyGen avatar SDK, and our own asynchronous post-processing pipeline.

What it does

Autonomous Market Research Agent – Pulls trending keywords, sentiment vectors, and competitor benchmarks from Twitter, Google Trends, and SEMrush APIs; ranks virality potential with a weighted scoring model (planned). Supernova helps creators understand why ideas go viral, by surfacing underlying trend + audience + format signals. We rank your idea’s growth potential in real-time before it’s even posted.
Personal Deep Profiling Engine – Uses LangChain agents to analyze YouTube and LinkedIn profile URLs, extracting insights on tone, vocabulary, and typical sentence structure to synthesize a comprehensive psychographic profile.
Script Generator (LangChain) – Using tools like market JSON + voice embeddings + notion of target duration. Outputs parallel A-Roll (spoken) and B-Roll (visual directions) with millisecond-level timestamps.
HeyGen Avatar Synthesizer – Sends cleaned A-Roll to HeyGen /v2/video/generate with user-selected avatar & voice; waits on a task-poller that backs off exponentially. Also, We go beyond demographics; your tone, structure, and sentence patterns train your own creator LLM.
B-Roll Retriever – Extracts keywords from B-Roll directions and searches Pexels video API for relevant clips.

How we built it

Layer	Tech	Highlights
Frontend	Next.js 13 / React 18, ShadCN UI, Tailwind CSS, Zustand	Suspense-enabled data fetching, granular skeletons, custom useProgress() hook for multi-stage modals.
LLM Orchestration	LangChain 0.2.x, OpenAI GPT-4o	ChainOfThought → ReAct agent with custom MarketSearchTool, UserVoiceTool, and TimestampValidator.
Avatar & Voice	HeyGen REST v2	Dynamic caption templates, vertical dimension override, instant avatar upload simulation in demo mode.
Post-Processing	Worker-Threads + fluent-ffmpeg + wasm-ffmpeg fallback for Vercel Edge (planned)	Streams, never writes to disk; uses -filter_complex [0:v][1:v]overlay for B-roll.

Challenges we ran into

Synchronizing B-roll & A-roll: Ensuring captions, avatar lip-sync, and overlay video aligned perfectly required timestamp validation, whisper-based alignment, and clip trimming.

Accomplishments that we're proud of

Built a creator-first video pipeline that goes from input idea → market intelligence → personalized avatar → finished video, all in a single workflow. Zero-camera, zero-editor.
Designed a modular LangChain agent stack that orchestrates multiple tools for market scanning, voice personalization, and timestamp alignment.

What we learned

Async orchestration across AI tools is hard but doable with a strong agent architecture.
AI content generation is about defaults. Users want fast, simple, optimized outputs—not endless sliders.

What's next for supernova

Auto-posting integration – push finished videos directly to TikTok/IG/YouTube via their publish APIs.
Multilingual dubbing – leverage HeyGen multi-voice + Whisper v3 to output 15 languages in parallel.
Brand style-guide ingestion – parse PDFs / Figma tokens → apply brand-fonts & colour palette automatically.
AI performance predictor – fine-tune a regression model on historical TikTok analytics to forecast views before publishing.
All-in-one creator hub – suggest optimal posting times, auto-repurpose content for blogs & newsletters.