itssujeeth Distinguished Architect / Technology Director · 2024 In production
Enterprise Model Delivery Platform
An MLOps and LLMOps platform for a Fortune 500 automotive aftermarket retailer, originated with a working demo, won in April 2025, approved by the architecture review board, and carrying 9 of 27 models in production by September 2026.
The data strategy division had 27 classical ML models, each built by a data scientist in a notebook on a workbench instance. Some had online endpoints, others ran predictions straight from the notebook, and results often went out as CSV files feeding dashboards. The organization developed models one at a time. Each team found its own way to deploy, and there was no standard route from a notebook to a monitored production endpoint, no CI/CD for models, and no shared practice for quality monitoring. Generative AI adoption was coming and would land on the same gap.
I identified the need, built a working demonstration of a model moving through a standard pipeline with CI/CD and quality monitoring, and worked with the commercial team to convert that demonstration into a funded platform engagement. I then led the platform architecture and its governance: CI/CD, repeatable model onboarding, deployment, and monitoring, with support for several ML frameworks so teams were not forced onto one. LLM lifecycle practices, including evaluation and operational governance, were designed in from the start.
Measured: as of September 2026, 9 of the 27 models were live on the platform, with the rest on a roadmap. The architecture passed the enterprise architecture review board. Other portfolios adopted the platform, and it extended to LLMOps as planned, including a promptfoo-based LLM evaluation framework and onboarding automation for new use cases.
This engagement started as a demonstration, not a request. The organization had capable data scientists and real models, but every one of them reached production by a different route, and most never reached monitored production at all. I built a small end-to-end demonstration that took a model through a standard pipeline with CI/CD and quality monitoring, showed it to the sponsor, and that was what unlocked the funding. The demonstration did the selling. The commercial team shaped the engagement around it.
The project was awarded in April 2025 but initially ran without a dedicated architecture lead, and it stalled for about six months on the CI/CD stack choice. When I joined in September 2025 I set the vision, principles, roadmap, and plan, produced the architecture and cost expectations, and took them through the architecture review board. What I would do differently is educate the sponsor on the process, not just the technology. The demo won the funding. It did not win the right team.
The design decision that mattered most was to build for the model lifecycle rather than a single use case. Repeatable onboarding, framework choice left to the team, deployment and monitoring as platform services, and an evaluation and governance path for LLMs planned in before anyone had asked for one. When generative AI adoption arrived, the platform already had a place for it, and the same practices now carry LLM evaluation.
Taking the architecture through the enterprise architecture review board was part of the work, not an afterthought. Governance approval is what lets a platform become the default path rather than one more option.
Client and employer details are confidential. The related Vertex AI Model Monitoring PoC is a public, synthetic-data exploration of the monitoring patterns this platform relies on.