• Built production LLM assistants on AWS: Bedrock + LangChain + RDS for PostgreSQL (pgvector) / OpenSearch for RAG, with hybrid retrieval + caching, achieving response time <2s.

• Implemented agent orchestration, evaluation, and safety: LangGraph workflows; Ragas/DeepEval gates in CI/CD; prompt injection and PII guardrails with Bedrock Guardrails.

• Fine-tuned domain LLMs on client data using Unsloth AI (QloRA, supervised fine tuning) on local GPUs and cloud GPU clusters, reducing prompt length for repeated queries by up to 40%.

• Built LLM -> Text-to-speech pipelines for long-form content generation with AWS Bedrock and

Amazon Polly, ffmpeg encoding for HLS streaming + Cloudfront caching, achieving content request to

stream start time <10s.

• Built NeptuneML + PyTorch (DGL) recommender lifting daily engagement by 50%; automated ML model retrains with Step Functions.

• Reduced p95 latency by 90% and DB cost by 85%: replaced MySQL monolith with Neptune + DynamoDB + Redis cache.

• Deployed Rekognition object-detection in UGC for recommender and moderation, reducing human review time by 70%.

• Automated video ingest (MediaConvert + Lambda), cutting video feed load-time by 90%.

• Implemented click-stream analytics: Kinesis to Timestream content leaderboards powering real-time

recommendations that increased daily post views by 30%.

• Rolled out CI/CD (GitLab, CodePipeline) and 100% CDK/CloudFormation across multi-account AWS org.

• Designed rate-limited GraphQL and REST APIs (API Gateway + OpenAPI).

• Integrated CloudWatch, X-Ray, OTEL for real-time observability.

• Created v1 of 2 web apps (React + Go), hired and led engineers, maintained quality through code reviews.

Built and maintained web and mobile apps for 10+ startups from multiple countries.