I take AI systems from model to production — end to end, on cloud.

As AI engineer

• Orchestrated six specialized LLM agents, running the three generation agents in parallel on Step Functions Express to cut end-to-end latency to 30–45 seconds.
• Routed models by task across Amazon Bedrock — Nova Lite for structured extraction, Claude Haiku 4.5 for writing — holding cost at ~$0.02 per application.
• Built server-side PDF parsing with pdfplumber, materially improving extraction accuracy on multi-column resumes over browser-based parsing.

As cloud / systems engineer

• Designed the serverless backbone — Step Functions, Lambda, API Gateway, DynamoDB — for under $3/month of infrastructure.
• Bypassed the API Gateway 29-second timeout with an async ingestion pipeline: a containerized Playwright Lambda on ECR with DynamoDB polling.
• Shipped the frontend on S3 + CloudFront with HTTPS (ACM), a custom domain on Route 53 and least-privilege IAM, following the Well-Architected Framework.

As AI engineer

• Trained a ResNet34 CNN on Amazon SageMaker to ~80% accuracy on image-to-genre classification.
• Paired Grad-CAM with GPT-4o so every prediction comes with a natural-language justification.

As cloud / SRE engineer

• Designed a cost-efficient inference architecture on EC2 Spot + Lambda + S3, cutting inference cost by ~90%.
• Delivered the Flask app for real-time upload, prediction and playlist recommendations.

As AI engineer

• Forecast pollution trends with SARIMAX at 85% accuracy, turning raw readings into proactive public-health alerts.

As data / systems engineer

• Built the ingestion path from Arduino and ESP32 sensors measuring CO, CO₂ and NH₃ into a live Flask + Plotly dashboard for urban-planning teams.