Vega AI

Inspiration

Small brands and solo founders spend 10+ hours a week on content alone — writing captions, designing visuals, tracking trends. Most can't afford a content team. We built Vega AI because we lived that problem ourselves while building in public. The idea was simple: what if your entire content operation ran on autopilot, and actually got smarter over time?

What it does

Vega AI is an autonomous content engine for small businesses and solo founders. You describe your brand once — tone, niche, audience, colors. Three AI agents handle everything else:

Brand Agent reads your Brand Brain and orchestrates the daily loop
Content Agent generates brand-specific images and videos using Wan2.2 Flash on A100 GPUs, enhanced by Groq (llama-3.3-70b)
Scout Agent monitors trends daily and feeds insights back into your Brand Brain

Generated content is delivered directly to your Discord or Telegram. You approve or reject. The system learns either way.

How we built it

Frontend: React + TypeScript, deployed on Vercel
Backend: Flask on AWS EC2 (eu-north-1), routed via Cloudflare Tunnel
Database: AWS DynamoDB for campaign records and all generated content metadata — chosen for its ability to scale to millions of items without provisioning or maintenance overhead
Auth & state: Firebase Auth + Firestore for user profiles and brand brain
Storage: AWS S3 for all generated images and videos
AI pipeline: Groq API for prompt enhancement and trend analysis, Modal (A100 40GB) for Wan2.2 Flash image and video generation
Delivery: Discord Bot + Telegram Bot for direct message delivery
Scheduling: Custom PM2 scheduler triggering daily at 9AM per user timezone
Agent orchestration: Sequential state machine — awaiting_checkin → checkin_done → awaiting_approval → approved → awaiting_scout → done

Challenges we ran into

Discord bots cannot DM users without a shared server — solved by building a dedicated Vega Discord server as a mandatory onboarding gate
Wan2.2 on Modal hitting OOM errors on A100 — fixed with enable_model_cpu_offload and lazy initialization
Month-end scheduler crash (day 31 + 1 = day 32 bug) — fixed with timedelta
Firestore gallery subcollection missing entries due to missing timezone import — silent failure caught through log analysis
Discord interaction timeout — bot was calling the agent before responding to Discord's 3-second deadline, causing "interaction failed" errors on every button press

Accomplishments that we're proud of

A fully autonomous three-agent loop that actually runs end to end in production
Brand Brain that gets smarter with every campaign through Scout Agent trend injection
Real AI-generated visuals delivered to users daily with zero manual input
Built and deployed a production stack — Vercel, AWS EC2, DynamoDB, S3, Modal — as a solo team in weeks
Gallery displaying real generated content inside the dashboard with download support

What we learned

DynamoDB's schemaless design is genuinely powerful for content metadata at scale — no migrations, no schema headaches, just write and query
Wan2.2 needs extremely specific motion and composition descriptors or it defaults to generic output — prompt engineering for video generation is its own discipline
Silent failures in async agent loops are the hardest bugs to catch — aggressive logging at every state transition saved us multiple times
Building for judges is different from building for users — the testing experience matters as much as the product itself