Echo

Echo is your brand's voice, amplified — it learns how you sound, tracks what's trending, and writes posts that feel like you.

Comment

Inspiration

Every brand on Instagram is flying blind. You post constantly, but you never really know two things: does this sound like us, and does it match what's actually working in our niche right now? Generic AI writers spit out off-brand text you have to rewrite. Schedulers help you post on time but never make the post better. We wanted something that learns your voice from your own history and measures it against live trends — so every post is both on-brand and on-trend.

What it does

Echo is an AI content engine for Instagram that runs in two modes:

  • Audit Mode — Import your past posts (as JSON). Echo learns your brand voice, then critiques your history against current trends: what you're missing, what's underperforming, and how to fix it.
  • Creation Mode — Give Echo a one-line product brief and it generates a brand-new, trend-optimized post, written in your established voice and chosen tone.

Optionally, you can feed it captions or images you admire and pick your genre to steer the style even further. The result is content that feels like you wrote it on your best day.

How we built it

  • Frontend: a mobile-first React + Vite + Tailwind PWA — installable, fullscreen, no app store.
  • Brand voice via "RAG-lite": instead of fine-tuning, we inject 2–4 of the user's real sample posts straight into the prompt, so the model mirrors their voice from tiny data.
  • Synthesis: a cloud LLM for text generation and trend analysis, plus an image-generation API for visuals.
  • Trend layer: the model measures the brand's content against a curated "trending now" signal for its niche.
  • Plumbing: Vercel serverless functions keep API keys server-side, localStorage persists the brand profile, and the app is a short linear flow — Brand Foundation → Inspiration → Prompt → Output.
  • Optional edge layer: an on-device WebLLM/WebGPU draft that runs privately on the phone itself.

Challenges we ran into

  • Voice from almost no data. Fine-tuning was off the table in a sprint, so we engineered the RAG-lite injection until a handful of samples reliably captured tone.
  • Making "trend match" concrete rather than hand-wavy — turning a fuzzy idea into a real, actionable audit.
  • Keeping the model honest — strict guardrails so it never fabricates specs, prices, or stats that aren't in the brief.
  • Never crashing on bad output — defensive JSON parsing with per-field fallbacks, so a malformed response degrades gracefully instead of breaking the screen.

What we learned

  • For brand voice on a hackathon timeline, RAG-lite beats fine-tuning — and it's faster and cheaper.
  • Auditing existing content is as valuable as generating new content — that insight reshaped the whole product into two modes.
  • Tight prompt engineering plus guardrails is what separates a demo that looks smart from output a creator would actually post.

What's next for Echo

  • Live trend-jacking from real-time feeds instead of a curated list.
  • A closed feedback loop — the voice model sharpens every time you approve or edit a post.
  • Auto-scheduling, plus expansion back into Reel scripts and X threads from the same single input.

Built With

Share this project:

Updates