Haru

Haru turns what you meant to say into what you should have said and then teaches you to not need it. An AI communication coach for global professionals, live with real payments.

Comment 2

Inspiration

Every professional knows the gap: you know exactly what you mean, but in a high-stakes moment like a client call, a salary negotiation, an escalation email which it comes out hedged, apologetic, or flat. For India's millions of client-facing professionals working in their second or third language, that gap costs deals, promotions, and confidence. I wanted to build something that closes it and that doesn't just rewrite your words, but actually teaches you to not need it.

What it does

Haru is an AI-powered executive communication coach with two modes:

  • Voice Mode — speak your unfiltered thoughts plus context ("tell the client we missed the deadline, keep it professional"). Haru transcribes, strips filler, and returns a polished script with tone analysis (confidence, clarity, diplomacy scored 0–100), alternative framings, and a side-by-side diff.
  • Text Mode — paste a draft, pick your audience and goal, and get a rewrite with inline grammar fixes and a one-line "how this will land" prediction.
  • The Learning Engine — every refinement logs your recurring weak spots and auto-generates personalized spaced-repetition flashcards (SM-2). You close your own communication gaps over time instead of memorizing generic vocab.

How I built it

  • Frontend: Next.js (scaffolded with Vercel v0, then completed in local development), deployed on Vercel at haru.co.in.
  • Database: Amazon Aurora PostgreSQL (Serverless v2), connected to Vercel in one click — users, sessions, mistake patterns, flashcards, and subscriptions in one relational schema, which the spaced-repetition engine depends on.
  • AI inference: Amazon Bedrock for the refinement, tone scoring, and framing.
  • Auth & email: started on Cognito + SES, later moved to Better Auth + Brevo (more on that below).
  • Payments: Razorpay with UPI Autopay — the right fit for the Indian market.

Challenges I ran into

  • SES production access never came. I requested it five days before the deadline and heard nothing. Cognito's email verification depends on SES and caps at 50 sandbox signups that land in spam — unshippable. After four days of waiting, I made the call to rip it out and move auth to Better Auth and email to Brevo (300 free emails/day, instant setup). Two hours, and the app was unblocked.
  • v0 credits burned fast. My workflow — build a feature, test thoroughly, fix bugs, repeat — used up the full v0 budget in two days. I pulled the code to GitHub, finished it locally with AI assistance, and kept deploying to the same Vercel project.
  • Razorpay onboarding. My existing account was registered for freelance work, so the SaaS site was rejected. I created a SaaS-categorized account and got approved in ten minutes.

What I learned

The biggest lesson was knowing when not to use a service. Aurora and Bedrock were the right calls and stayed. Cognito and SES were good services gated by an approval process slower than a hackathon window — so the pragmatic move was to swap them for tools that ship at startup velocity. v0 is excellent for scaffolding; local development with AI assistance is better for the connected, debug-heavy work. Use each where it's strongest.

What's next for Haru

On-device voice transcription (via in-browser Whisper) for privacy and cross-browser consistency, a Chrome/Gmail/Slack extension to refine text where people actually write, and a mobile app with on-device transcription and synced learning progress.

Share this project:

Updates