Professional Experience

Situation: Kinetix's GPU costs for Whisper inference were unsustainable at scale. Their NVIDIA-based infrastructure threatened unit economics as customer volume grew.

Task: Evaluate AWS Inferentia as a cost-effective alternative and migrate the Whisper inference stack without compromising transcription quality or latency.

Action: I ported the Whisper model to AWS Inferentia using the Neuron SDK, conducted rigorous benchmarks against NVIDIA Triton across various audio workloads, and implemented KV-Cache and I/O optimizations.

Result: Achieved 50% reduction in inference costs with a 10% performance improvement. Delivered a comprehensive benchmark report that informed Kinetix's long-term infrastructure strategy.

Situation: The application was tightly coupled to OpenAI, creating vendor lock-in. Additionally, valuable enterprise knowledge was trapped in SharePoint repositories.

Task: Enable multi-provider LLM flexibility without disrupting existing functionality, and build a RAG pipeline to unlock SharePoint knowledge.

Action: I architected a feature-flip system for seamless OpenAI/Bedrock switching. Built a SharePoint crawling mechanism, indexed documents in OpenSearch with semantic chunking, and integrated a Bedrock agent with knowledge base.

Result: Delivered multi-provider LLM flexibility allowing runtime provider switching, plus a fully operational enterprise RAG system with secure, permission-aware document retrieval.

Situation: The team had no industrialized way to measure LLM translation quality. Manual spot-checks couldn't scale, and regressions were only discovered after impacting customers.

Task: Build an automated LLM evaluation pipeline integrated with CI/CD, and implement observability for real-time monitoring of usage, latency, and costs.

Action: I implemented a DeepEval-based evaluation framework directly in the CI/CD pipeline. Added Sentry tracing for LLM calls and integrated Comet translation quality metrics.

Result: Established automated CI/CD LLM testing that catches regressions before deployment. Enabled live usage and cost monitoring giving leadership full visibility into AI spend.

Situation: AWS deprecated their Forecast service, forcing an urgent migration. Bevolta's system provided per-client forecasting models critical for personalization.

Task: Migrate from AWS Forecast to SageMaker AutoML while preserving per-client model personalization and implementing intelligent retraining.

Action: I designed a migration architecture using Autopilot with multiple algorithms. Built a DynamoDB-based performance tracking system with conditional retraining triggers. Created two operational modes: cost-conscious and exploratory.

Result: Achieved functional migration with zero service disruption. Delivered intelligent performance monitoring that triggers retraining only when needed, plus FinOps optimization reducing ongoing ML infrastructure costs.