An AI conversational companion that helps you decide what to gift — not just what to buy.
Picking a thoughtful gift is one of the highest-frequency, highest-emotional-stakes decisions in everyday life. Despite hundreds of e-commerce platforms, gift-giving stays stressful: shoppers feel decision fatigue, default to repeat purchases, or settle for cash that feels impersonal.
The friction isn't "I don't know where to buy." It's "I don't know what to give."
Existing tools fail this:
- Marketplaces (Amazon, Flipkart) surface 10,000 SKUs and offload the decision back to the user.
- Generic AI chatbots (ChatGPT, Gemini) dump ten-item lists with no purchase path.
- Niche curators (FlowerAura, Ferns N Petals) work for one category but can't navigate context.
Dearly replaces browse and pick with think it through together.
Users describe a situation — who they're gifting, the occasion, the relationship, the budget — and Dearly's AI walks them through a structured conversation, asks the right follow-up questions, then converges on 2–3 specific recommendations grounded in real Google Shopping India results, complete with reasoning ("why this works", "the risk", "how to make it better").
It's intentionally narrow: not a shopping engine, not a wishlist app, not a generic AI chat. A decision companion.
Try it now → dearly-beryl.vercel.app
You can use Try It Free without signing up. Click around. The product paywalls after the first recommendation — sign up free to continue.
Drop screenshots into
docs/screenshots/and the images below will render. See "How to add screenshots" below.
| Landing | Conversation |
 |
 |
| Event Dashboard | Trial Paywall |
 |
 |
Structured 4-stage conversation flow (Opening → Mode Choice → Quick/Thoughtful → Execution) implemented as a Groq system-prompt contract. The AI returns JSON with reply, choices, search_queries, and max_price_inr — never an unbounded paragraph of ideas.
SerpAPI Google Shopping (India, INR) fans out one query per gift concept. A budget-tiered filter rejects mismatched prices (no ₹150 trinkets recommended on a ₹15,000 brief) and sorts by descending value-in-range.
Each gifting situation is a persistent thread. Users can manage Mom's Birthday, Anniversary, and Boss's Promotion as three separate events, each with its own cover image, status badge, and conversation history.
Upload a cover image, drag to reposition, edit the title inline, watch the status update — every interaction is persisted to Supabase with proper RLS.
Free users get one full advisory conversation. The paywall blurs the chat and gates further interaction — clean wedge between curiosity and commitment.
Sidebar collapses into a drawer. Hero collage scales. Event bar scrolls horizontally. Tested on iOS Safari, Chrome Android, and desktop Chrome / Firefox / Safari.
Supabase email + password with proper SSR middleware. Sessions refresh transparently; protected /chat route redirects unauthenticated users to login.
| Layer | Choice | Why |
|---|---|---|
| Framework | Next.js 16 (App Router) | SSR for SEO; single repo for front-end + API routes; Vercel-native |
| UI | React 19 + Tailwind CSS 4 | Fast iteration, no design-system overhead at MVP scale |
| LLM | Groq (Llama 3.3 70B) | Sub-second token latency, ~95% cheaper than GPT-4 at comparable quality for this task |
| Auth + DB + Storage | Supabase (Postgres) | One vendor for the four hardest infra problems; RLS for free; managed Postgres |
| Product Retrieval | SerpAPI Google Shopping (gl=in) | Live data, India-localized, no scraping risk |
| Hosting | Vercel | Zero-config Next.js, instant rollback, preview branches |
| CI/CD | GitHub → Vercel auto-deploy | Push to main → live in ~60 seconds |
┌─────────────────────────────────────────────────────────────────┐
│ User (Browser) │
└────────────────────────────┬────────────────────────────────────┘
│ HTTPS
┌────────────────────────────▼────────────────────────────────────┐
│ Next.js on Vercel │
│ ┌────────────────────┐ ┌──────────────────────────────────┐ │
│ │ React (App Router) │ │ API Route: /api/chat │ │
│ │ - Landing │ │ 1. Groq LLM (JSON contract) │ │
│ │ - Chat UI │◄──►│ 2. SerpAPI fan-out (parallel) │ │
│ │ - Modals │ │ 3. Budget-tier filter │ │
│ └────────────────────┘ │ 4. Compose reply + products │ │
│ │ └──────────────┬───────────────────┘ │
└─────────────┼─────────────────────────────┼─────────────────────┘
│ │
▼ ▼
┌──────────────────────┐ ┌────────────────────────────┐
│ Supabase │ │ External APIs │
│ - Auth (cookies) │ │ - Groq /chat/completions │
│ - Postgres + RLS │ │ - SerpAPI google_shopping│
│ - Storage (images) │ │ │
└──────────────────────┘ └────────────────────────────┘
| Stage | p50 | p95 |
|---|---|---|
| Network round-trip | 120 ms | 300 ms |
| Groq inference (Llama 3.3 70B) | 800 ms | 1.6 s |
| SerpAPI batch (3 parallel) | 1.2 s | 3.0 s |
| Server composition | 30 ms | 80 ms |
| Client render | 20 ms | 60 ms |
| End-to-end perceived | 2.2 s | 5.0 s |
dearly/
├── app/
│ ├── api/chat/route.js # LLM + SerpAPI orchestration
│ ├── auth/callback/route.js # Supabase OAuth return URL
│ ├── chat/page.js # Main authenticated dashboard
│ └── page.js # Landing + login/signup modal
├── components/
│ └── ProfileModal.js # Avatar + display name editor
├── lib/supabase/
│ ├── client.js # Browser client
│ ├── server.js # Server-side client (cookies)
│ └── middleware.js # Session refresh helper
├── public/ # Robot avatars, hero illustrations
├── proxy.js # Next.js 16 middleware (auth gate)
├── next.config.mjs # Image domains, lint config
└── Dearly-PRD.docx # 30-page product requirements doc
# 1. Clone the repo
git clone https://github.com/GulshaPanda/dearly.git
cd dearly
# 2. Install dependencies
npm install
# 3. Set up environment variables
cp .env.example .env.local
# Then fill in the four keys (see .env.example)
# 4. Run the dev server
npm run dev
# 5. Open http://localhost:3000| Variable | Get it from |
|---|---|
GROQ_API_KEY |
console.groq.com — free tier is plenty |
SERPAPI_KEY |
serpapi.com — 100 free searches / month |
NEXT_PUBLIC_SUPABASE_URL |
supabase.com → New project |
NEXT_PUBLIC_SUPABASE_ANON_KEY |
Same Supabase project, Settings → API |
After creating a Supabase project, run the SQL in docs/schema.sql to create the events and messages tables, enable RLS, and create the avatars storage bucket.
The full product requirements document is committed in this repo:
Dearly-PRD.docx — 19-section PRD covering problem framing, market sizing, personas, competitive analysis, functional requirements, data model, monetization strategy, GTM plan, risks, and 12-month roadmap. Written at the level of a senior PM at a top tech company.
V1 (shipped) — Conversational AI flow, product retrieval, multi-event memory, trial paywall, mobile-responsive.
V1.5 (next 90 days) — Auto-extract event metadata from chat, calendar reminders via email, forgot-password, Google OAuth, PostHog analytics.
V2 (3–6 months) — Pro subscription (₹199/mo), Surprise Delivery composer, recipient profiles with full gift history, Product Hunt launch.
V3 (6–12 months) — Group gifting with UPI split, curated small-business marketplace (affiliate revenue), WhatsApp bot, Hindi localization.
Full roadmap with KPIs and OKRs in the PRD.
Honest reflection from someone who actually shipped this:
- Event auto-extraction was a missed V1 unlock. Today users edit the title manually. The AI already knows it's "Mom's Birthday" by message two — should have piped that into the events table from day one.
- The paywall is too hard for the test cohort. Smart bet for monetization signal, but probably costs me early word-of-mouth. I'd A/B a softer 2nd-recommendation paywall in V1.5.
- No analytics in V1. I'm flying blind on the trial → signup funnel. PostHog is the first V1.5 task.
- Email confirmation off in production. A velocity call to ship faster; before scaling, this gets flipped on with a proper transactional email service (Resend).
Gulsha Panda — building products for the Indian consumer market.
I designed, scoped, built, deployed, and shipped this end-to-end as a portfolio project for APM applications. The repo includes the PRD, the codebase, the live product, and this README — together they're meant to demonstrate the full PM-meets-builder loop, not just spec-writing.
- Email: gulshapanda@yahoo.com
- LinkedIn: (add your link)
- Twitter / X: (add your link)
MIT — see LICENSE.
If you found this interesting, star the repo — it helps a lot.