DriftSpot 🌊

Hyperlocal travel discovery for time-constrained travelers.

"I was at a conference in Whitefield, Bangalore. It was 6 PM. I had the evening free but no idea what was nearby — Google Maps showed places 45 minutes away in traffic, TripAdvisor gave me a city-wide Top 10. I ended up staying at the hotel. I built DriftSpot so that never happens again."

Live Demo · API Docs · Product Docs

▶ Watch Demo on Loom

---

What It Does

DriftSpot takes a single free-form query and returns 2–3 curated, opinionated place recommendations — filtered for where you actually are, what time it is, and how you feel right now.

"I'm done with my conference in Whitefield. It's 6 PM, I want to relax somewhere quiet nearby."

→  Windmills Craftworks Brewery   ·  12 min  ·  "mellow craft beer evening"
→  Cubbon Park                    ·  18 min  ·  "quiet and green, good for a walk"
→  Koshy's                        ·  20 min  ·  "old-world Bangalore, filter coffee"

No search. No filters. No scrolling. Just go.

---

Architecture

User Query (Natural Language)
          │
          ▼
┌─────────────────────────────────────────┐
│           FastAPI Backend               │
│                                         │
│  JWT Auth (Supabase)                    │
│          │                              │
│          ▼                              │
│   ┌─────────────────┐                   │
│   │  ExtractIntent  │  GPT-4.1-mini     │
│   │                 │  Expert + Few-shot│
│   └────────┬────────┘  prompting        │
│            │ intent {mood, purpose ...} │
│            ▼                            │
│   ┌─────────────────┐                   │
│   │  DiscoverPlaces │  Google Places    │
│   │                 │  Nearby Search    │
│   └────────┬────────┘  keyword-filtered │
│            │ candidate places[]         │
│            ▼                            │
│   ┌─────────────────┐                   │
│   │   RankPlaces    │  GPT-4.1-mini     │
│   │                 │  Strict rules,    │
│   └────────┬────────┘  no forced fits   │
│            │ recommendations[]          │
└────────────┼────────────────────────────┘
             │
             ▼
    Supabase (PostgreSQL + RLS)
    queries · saved_places · feedback

The pipeline is built with LangGraph — each stage is a discrete, independently testable node with typed state flowing between them.

---

Key Technical Decisions

1. LangGraph for AI Pipeline Orchestration

Rather than chaining everything into a single monolithic prompt, the pipeline is broken into three nodes with a shared TravelState TypedDict. This makes each step independently debuggable, swappable, and testable — critical for a system where prompt changes in one step affect the entire output chain.

2. purpose Field for Targeted Place Search

Instead of hardcoding a mood → place-type mapping, the LLM extracts the traveler's specific activity goal as a Google Places keyword:

"I want to smoke somewhere" → purpose: "hookah bar OR tobacco shop"
"eat biryani"              → purpose: "biryani restaurant"

This keyword is passed directly to Google Places — so discovery is targeted before ranking, not after.

3. JWT Auth + Supabase RLS (No Service Role Key on Client)

B2C auth via Google OAuth. The frontend only holds the user's JWT — no admin keys ever leave the backend. Row Level Security enforces per-user data isolation at the database layer:

create policy "Users access own data"
  on saved_places for all using (auth.uid() = user_id);

Even a compromised client cannot read or modify another user's data.

4. Persistent Schema Designed for Future Graph Recommendations

Every query, shortlist, and feedback action is stored with structured fields (not raw blobs), designed to be projected into a user–intent–place graph:

(User) -[:SEARCHED_IN]-> (Location)
(User) -[:SHORTLISTED {vibe, travel_time}]-> (Place)
(User) -[:LIKED / :DISLIKED]-> (Place)

This enables collaborative filtering and preference learning from Day 1 — without a painful schema migration later.

5. Expert + Few-Shot Prompt Engineering

Two techniques in combination:

• Expert persona — "You are a knowledgeable local travel guide..." activates domain-specific reasoning embedded in the model's weights
• Few-shot examples — constrain output vocabulary to values the downstream system expects ("relaxed" not "chill")
• Strict rules layer — prevents creative reinterpretation ("If someone asks to smoke, do NOT recommend salons")

---

Tech Stack

Layer	Choice	Why
Backend	FastAPI	Async-native, auto OpenAPI docs, Pydantic validation
AI Orchestration	LangGraph	Stateful multi-node pipelines; independently testable nodes
LLM	GPT-4.1-mini	Best quality/cost ratio for structured extraction + ranking
Place Data	Google Places API	Largest real-time coverage; open/closed status; keyword search
Auth + DB	Supabase	Google OAuth; RLS; PostgreSQL structured for graph-readiness
Frontend	React + Vite + Tailwind	Fast iteration; warm amber/orange design system
Hosting	Render (API) + Vercel (web)	Zero-config deploys from GitHub

---

Product Documentation

Document	Description
PRD	Problem statement, target user, north star scenario
User Stories	Acceptance criteria for V1 and V2 features
Sprint Plan V1	5-day build plan for the core discovery loop
Sprint Plan V2	Trust signals, cab booking, feedback loop
Architecture	ADRs — the why behind each technical decision
Startup Foundation	Market sizing, business model, GTM, competitive analysis

---

Running Locally

Backend

cd apps/api
pip install -r requirements.txt
cp .env.example .env   # fill in your keys
uvicorn main:app --app-dir src --reload
# API at http://localhost:8000  ·  Docs at http://localhost:8000/docs

Frontend

cd apps/web
npm install
cp .env.example .env   # fill in your keys
npm run dev
# App at http://localhost:5173

Required environment variables

Variable	Where to get it
OPENAI_API_KEY	platform.openai.com
GOOGLE_PLACES_API_KEY	Google Cloud Console → Maps → Places API
SUPABASE_URL	Supabase Dashboard → Project Settings → API
SUPABASE_ANON_KEY	Supabase Dashboard → Project Settings → API
VITE_SUPABASE_URL	Same as SUPABASE_URL
VITE_SUPABASE_KEY	Same as SUPABASE_ANON_KEY
VITE_API_URL	http://localhost:8000 (local) or your Render URL

---

Roadmap

• Natural language intent extraction (mood, purpose, constraints)
• Hyperlocal discovery via Google Places with keyword filtering
• LLM-ranked recommendations with context-aware explanations
• Google OAuth + JWT auth
• Saved places (shortlist) persisted to Supabase
• Left panel showing shortlisted spots across sessions
• Trust signals per recommendation (safety, crowd, solo-friendliness)
• Cab booking deeplink (Uber/Ola) from recommendation card
• Thumbs up/down feedback loop for personalization
• Graph-based collaborative recommendations (Neo4j)
• GPS auto-location

---

About

Built by Aditya Dhir — Senior AI Engineer.

DriftSpot demonstrates the intersection of AI engineering (LangGraph, prompt engineering, LLM orchestration) and product thinking (PRD, sprint planning, competitive analysis, investor framing).