At Leximind LLC, I built an internal knowledge assistant with LlamaIndex and LangGraph on Azure OpenAI and Amazon Bedrock. It serves more than 1,500 employees, with P95 response time under two seconds.
I own SageMaker Pipelines for my team's main models, from training and tuning through model registry and inference. I cut P95 inference latency from 180ms to 65ms with TensorRT and quantisation, lowering GPU serving cost by 30%.
At Mission, I built FastAPI inference services handling over 500,000 requests a day and pipelines for OCR, transcription, speaker diarisation, and image tagging. A PyTorch recommendation model I trained raised content click-through rate by 12%.
At Amperor, I built demand forecasting models that reduced forecast error by 18% and developed models for healthcare operations under HIPAA requirements. That early work still shapes how I approach data security.
