Kenneth Koski

B2B SaaS AI products, model evaluation, and production infrastructure

• Built ChatCPG, a customer-facing enterprise-data assistant combining structured PostgreSQL data with Weaviate vector search and retrieval-augmented generation (RAG) over uploaded documents. Implemented chunking, reranking, retrieval evaluation, and document-access controls; integrated LangChain and MCP tool use.
• Architected data and ML pipelines for RiskWise, a hosted emerging-risk analytics product; fine-tuned open models to reduce inference costs and selected Qwen 3.5 9B after comparisons with larger models showed sufficient task quality.
• Built and own an NLU system replacing manual vendor-data classification for a multibillion-dollar fintech company. Built agentic LLM evaluation pipelines for RiskWise and NLU with F1 acceptance checks (0.95 threshold; >0.99 test result) and human approval for deployment.
• Sole developer and operator of ix, an internal GPU fine-tuning and inference platform handling tens of thousands of multistage prediction runs, each starting with millions of input rows; integrated vLLM and versioned FastAPI endpoints, provider-failure recovery with idempotent completion checks, and granular cost tracking.
• Deployed and maintained Kubernetes and Argo Workflows; adapted Airflow workflows and automated deployments with Pulumi. Built incremental Snowflake-to-PostgreSQL sync that ran 10× faster through S3 exports and PostgreSQL bulk loading. Implemented tool permissions based on Auth0 groups and customer-data isolation.
• Align a four-person engineering team and five-person data science team around production model work; mentor where the disciplines meet and work with product stakeholders to translate requirements into platform decisions.

Open-source ML platforms, deployment automation, and integration testing

• Led Python integration architecture for Charmed Kubeflow, an open-source Kubernetes ML platform; built and maintained operators connecting upstream ML services with reproducible deployment and shared authentication.
• Developed a Juju charm for MLflow, packaging it for Kubernetes deployment.
• Built TensorFlow Serving deployment automation and a separate end-to-end pipeline covering model training, serialization, service readiness, inference, and prediction-accuracy checks.
• Automated Python/pytest and GitHub Actions validation in disposable AWS/MicroK8s environments; tested pipeline execution and model tuning and adapted integrations to upstream API changes.
• Maintained public integrations as a code owner, reviewing contributions and coordinating changes with users and upstream developers; built a Python deployment CLI with readiness checks and actionable errors.

• Redesigned the CMS Quality Payment Program's overnormalized PostgreSQL schema around actual access patterns, removing unnecessary joins; combined with parallel ETL processing, improved throughput 10×.
• Led two infrastructure engineers and built reproducible AWS GovCloud environments with Terraform, Ansible, and Jenkins; operated production services with health checks, centralized logs, alarms, and PagerDuty on-call.

• Designed a Python/OpenStack service for provisioning Lustre and extensible distributed storage on bare-metal HPC systems in an internal development project; provided architecture guidance, reviews, and mentoring.

• Owned application code, APIs, data, and production operations for a high-volume messaging product; improved PostgreSQL concurrency and asynchronous Celery/Redis processing as traffic grew.