FlowZint AI Hackathon 2026 — Entry

AI Support forModern E-Commerce

Combines RAG knowledge retrieval, real-time sentiment detection, and 11-language support to deliver instant, accurate, empathetic support — 24/7.

RAG knowledge retrieval
Sentiment detection
11 languages, auto-detect
Auto-escalation to agents
Streaming responses
Live analytics dashboard
ShopAssist · FlowMart
11
Languages
24/7
Availability
< 2s
Response time
RAG
Knowledge retrieval
15-step
Build plan
Free
Gemini tier

Built with

Next.js 14
Gemini 2.5 Flash
Supabase
Tailwind CSS
TypeScript
Recharts
Vercel
Features

Everything e-commerce support needs

Six capabilities engineered to make every customer interaction faster, smarter, and more human — without a human.

RAG Knowledge Retrieval

30% score

24+ FAQ chunks parsed, keyword-scored, and filtered. Only the top-5 sections relevant to each query are injected — grounded answers, zero hallucinations.

Sentiment Detection

30% score

130+ keyword dictionary across 4 tiers. ALL CAPS amplification, conversation context weighting, 30 hard escalation triggers. Zero latency.

11-Language Support

30% score

Unicode script-range detection plus Latin word patterns. Full RTL flip for Arabic. Translated UI in 11 languages. Gemini replies in whatever language you type.

Smart Auto-Escalation

Pulsing red banner with countdown timer, call button, and email transcript form. Fires automatically on anger, legal threats, or chargeback mentions.

Contextual Suggestions

Three category-colour-coded reply chips after every response. Staggered entrance animation. Keyboard navigation. Skeleton loading while streaming.

Analytics Dashboard

Live Recharts: daily volume, sentiment donut, language bars, top issues, escalation queue. Supabase-backed with realistic seed data for instant demos.

How it works

Message → answer in under 2 seconds

01

Customer sends message

In any language, any frustration level. Language is auto-detected via Unicode and sentiment is scored — before a single token hits Gemini.

02

RAG retrieves context

Knowledge base is chunked and scored against the query. Top-5 relevant sections are injected. System prompt gets emotional context for angry users.

03

Streamed clean response

Gemini streams back. Server buffers, strips SUGGESTIONS and [ESCALATE] markers, then re-streams clean text with a natural word-by-word delay.

Model Innovation — 30% of score

Three layers of AI intelligence

RAG, sentiment analysis, and multilingual NLP — each independently novel, together forming a cohesive intelligent support system.

RAG Pipeline

Grounded, accurate answers

24+ chunks parsed from store FAQ
Title (×5) / keyword (×3) / content (×1) weighting
Top-5 chunks per query, 3 000-token budget
ALWAYS_INCLUDE store info section
Force-reload endpoint for KB updates
Sentiment Engine

Emotion-aware responses

130+ keywords across 4 tiers (angry / negative / positive)
ALL CAPS detection: 1.6× score amplification
Repeated words + !! / ?? pattern detection
30 hard escalation phrases (legal, chargeback…)
Prior-turn context pulls score negative
Multilingual NLP

11 languages, zero config

Unicode script ranges: Hindi, Arabic, CJK, Korean…
Latin-script word patterns for 7 languages
dir="rtl" layout flip for Arabic
176 UI strings (16 keys × 11 languages)
Auto-detect ⚡ / manual override with language selector
Technical Architecture — 25% of score

Clean, production-grade patterns throughout

Type-safe end-to-end

Full TypeScript — Message, SentimentScore, AnalyticsData interfaces shared across client, server, and DB.

SSE streaming

Server buffers full Gemini response, parses metadata, re-streams clean content word-by-word at 14ms/token.

Graceful degradation

Every Supabase call returns null on failure. App runs perfectly in memory-only mode during demos.

Sliding context window

30-turn / 6K-token cap with safe MIN_RECENT_TURNS guarantee. trimHistory() keeps the latest exchange always.

Try ShopAssist right now

Ask about an order, write in Hindi, trigger an escalation — see every feature working live in under a minute.

Admin password: shopassist2026