Four rooms, four different ways to be wrong.

LLM and NLP systems at an NLP data platform — the domain I'd been aiming at since the first RAG pipeline I built for someone else's business. Two months in, which is the honest number, and the reason the chat on the right has an answer ready for that question.

• LLM systems
• NLP
• Production

• Architected an end-to-end computer-vision active learning monorepo — 8 integrated microservices under Docker Compose and the NVIDIA Container Toolkit, covering YOLOv7 automated labelling, continuous training and dataset QA.
• Built a high-throughput video sampling engine with OpenCV and FFmpeg, pulling critical frames straight from MediaMTX RTSP live streams to feed continuous evaluation loops.
• Deployed GPU-accelerated inference endpoints with automated dataset sync to cloud storage — decoupled MLOps, not a notebook.

What it changed: I stopped thinking of a dataset as a thing you have, and started treating it as a loop you run.

• Built a conversational commerce agent on AWS Bedrock (Nova Lite) with the Strands Agents framework — RAG retrieval through Bedrock Knowledge Bases, cross-session memory, live order-trend analysis, and MCP tool-use orchestration in a 5-iteration agent loop. Customers place orders in natural language over WebSocket.
• Automated an enterprise timesheet system: AppSync GraphQL, 12 DynamoDB tables, 5 EventBridge-scheduled Lambdas handling provisioning, SES reminders, submission enforcement, YTD chargeability via DynamoDB Streams, and biweekly archival.
• Designed a multi-tenant serverless data platform in AWS CDK — 11 decoupled micro-stacks stitched through SSM Parameter Store, with OpenSearch indexing, SQS FIFO queues and Step Functions for third-party payment orchestration.
• Shipped production-grade security: Cognito RBAC, API Gateway authorizers, Bedrock Guardrails, input sanitisation, per-user rate limiting — zero critical findings in a third-party penetration test.

What it changed: Spec-driven development. Writing the requirements and design docs before the code felt slow for about a week, then stopped costing me rewrites entirely.

A year of finishing things. Three products, each one a different lesson in how much of ML engineering is actually latency, state and interface.