Culinary AI: Smart Food Recommendation Chatbot

Technical Highlights

Full-Stack AI Engineering: End-to-end development dari data pipeline (Instagram scraping, video transcription) hingga production deployment dengan Docker dan cloud infrastructure.

Advanced RAG Architecture: Multi-stage retrieval system dengan semantic search, context-aware filtering, dan fallback mechanisms untuk ensure 100% response rate.

ML/AI Integration: Whisper AI untuk video transcription (500+ videos), AWS Bedrock untuk LLM inference, dan Qdrant vector database untuk semantic search.

Production-Ready: Automated data pipeline, cookie rotation untuk rate limit handling, error recovery, dan monitoring.

Skills Demonstrated

AI/ML Engineering

• ▸RAG Implementation: LangChain orchestration dengan custom retrieval logic
• ▸Vector Database: Qdrant setup, indexing strategy, dan query optimization
• ▸LLM Integration: AWS Bedrock (Claude 3.5 Sonnet) dengan streaming responses
• ▸Audio Processing: Whisper AI untuk transcribe 500+ Instagram videos
• ▸Prompt Engineering: Context-aware prompts dengan dynamic variables

Backend Development

• ▸FastAPI: Async endpoints, SSE streaming, error handling
• ▸Data Pipeline: Automated extraction dari 806 columns → 18 relevant features
• ▸Rate Limit Handling: Cookie rotation mechanism dengan exponential backoff
• ▸Database Design: Metadata schema untuk efficient filtering

Frontend Development

• ▸Next.js 14: App Router, Server Components, streaming UI
• ▸TypeScript: Type-safe development dengan Zod validation
• ▸Real-time Updates: SSE integration untuk streaming responses
• ▸Rich UI: Interactive cards dengan Instagram/Maps links

DevOps & Infrastructure

• ▸Docker: Containerization untuk consistent deployment
• ▸Cloud Deployment: Vercel (frontend) + Cloud Run (backend)
• ▸Environment Management: Multi-environment configuration
• ▸Monitoring: Logging, error tracking, performance metrics

Architecture & Design Decisions

Multi-Stage Retrieval: Implemented fallback mechanism (strict filter → relaxed filter → general search) untuk ensure always ada hasil relevan. Ini critical untuk UX.

Context-Aware Filtering: Automatic time-based categorization (sarapan/makan_siang/nongkrong/makan_malam) berdasarkan current time atau query keywords. No manual filter needed.

Rich Metadata: Structured response dengan Instagram links, Google Maps links, menu, harga, jam operasional. Bukan cuma text recommendation.

Cookie Rotation: Implemented smart rotation dengan 3+ cookie files untuk avoid Instagram rate limits saat download 500+ videos. Include random delays dan retry logic.

Streaming Response: SSE implementation untuk better perceived performance. Response time turun dari 5-10s jadi under 2s (perceived).

Data Engineering

Challenge: 637 Instagram posts dengan 806 columns, inconsistent format, info tersebar di caption/video/hashtags.

Solution:

• ▸Automated extraction dengan regex patterns (location, hours, menu)
• ▸Video transcription dengan Whisper AI (78% success rate)
• ▸Data validation dan cleaning pipeline
• ▸Structured metadata untuk efficient retrieval

Results: 806 columns → 18 relevant features, 500+ videos transcribed, ready untuk production RAG system.

Impact & Metrics

• ▸Response Time: Under 2 seconds (perceived latency dengan streaming)
• ▸Accuracy: 100% grounded responses (no hallucination)
• ▸Coverage: 500+ tempat makan di Samarinda dengan rich metadata
• ▸Transcription: 78% video transcription success rate (500/637 videos)
• ▸Extraction: 70% location, 50% hours, 94% hashtags extraction success

Technical Challenges Solved

Instagram Rate Limits: Implemented cookie rotation dengan 3+ accounts, random delays, dan retry logic untuk download 500+ videos tanpa permanent block.

Inconsistent Data: Built robust extraction pipeline dengan regex patterns, fallback logic, dan manual validation untuk handle inconsistent Instagram captions.

Context Window Optimization: Multi-stage retrieval untuk balance antara relevance dan coverage. Strict filter first, fallback to general search if needed.

Real-time Streaming: SSE implementation dengan proper error handling, connection management, dan graceful degradation.

Production Deployment: Docker containerization, environment management, monitoring setup, dan cost optimization untuk AWS Bedrock usage.

Read Full Story: Blog Post