Technical overview
How I build.
A closer look at the systems I’ve built — architecture, backend, data, deployment and UI — for engineers and technical leaders.
01 / Tech lead · 2026 — Present
AI Control Plane
I lead the build of State Street’s AI Control Plane: a single platform for the full lifecycle of an AI application — onboarding, governance, observability, monitoring and evaluation — serving 150+ applications.
Problem
AI applications reached production through manual approvals and email threads. There was no single place that tied an application to its owner, its policies, its traces and its spend.
- Python
- FastAPI
- React
- Docker
- Helm
- Azure AKS
Application lifecycle
Architecture
- Defined the lifecycle workflow from onboarding to production, with governance and audit reviews built into the platform.
- Backend services in Python and FastAPI; React frontend, built with AI coding assistants.
- Containerised with Docker and deployed as application-as-code through Helm charts on Azure Kubernetes Service.
- Integrated with the observability stack, so every application carries its own traces, usage and cost.
- Role-based views and charge-back models shaped with governance, audit and FinOps stakeholders.
Result
One control plane for 150+ applications. Owners see usage, cost and policy breaches; developers see traces; FinOps and governance get charge-back, policy and audit in one place.
02 / Architect · 2025 — Present
Enterprise LLM observability & evaluation
I architected the enterprise LLM observability platform that became the standard monitoring layer for 150+ production AI applications, including LLM agents, RAG pipelines and Responsible AI workloads. It processes about 4 million traces a day in near real time.
Problem
Teams were shipping LLM applications with no consistent way to trace requests, measure quality or track cost. In a regulated bank, that meant incidents were slow to diagnose, token spend was invisible, and risk teams had no audit trail.
- OpenTelemetry
- OpenInference
- Arize AI
- Python
- Azure Databricks
- Locust
Architecture
- Standardised instrumentation on OpenTelemetry / OpenInference, with custom Python SDK wrappers over Arize AI so teams could onboard with minimal code changes.
- Traces feed FinOps, operational and Enterprise Risk Management dashboards covering latency, token cost, error rates and compliance metrics.
- An LLM-as-Judge evaluation framework runs controlled A/B tests of prompt and pipeline changes against gold datasets, flagging hallucinations, regressions and toxicity.
- Redesigned distributed telemetry to eliminate multi-hop trace loss, and validated zero-drop behaviour under peak load with Locust.
- Medallion (Bronze → Silver → Gold) pipeline on Azure Databricks for trace data.
Result
One standard monitoring layer across the AI portfolio, faster incident diagnosis, and cost and risk visibility for leadership.
03 / Proof of concept · 2026
Ask SS — agentic alert triage
An agentic assistant MVP that automates triage of alerts from an internal alert monitoring system.
Problem
Resolving an alert meant an engineer manually tracing the root cause and searching SOPs for the right resolution — slow, repetitive work.
- AWS AgentCore
- Python
- AWS models
Architecture
- Agent built on AWS AgentCore.
- Analyses the relevant SOPs to identify the root cause and recommend a resolution.
- Human-in-the-loop approval step before any action is taken.
Result
Reduced alert closure time by 50%.
04 / Engineer · 2022 — 2025
RAG-as-a-Service and Responsible AI
Core components of State Street’s internal RAG-as-a-Service platform, plus Responsible AI reporting.
Problem
Teams building AI applications each needed reliable retrieval over their documents and data, and naive similarity search alone didn’t return relevant enough results.
- PostgreSQL / pgvector
- FastAPI
- Azure Databricks
- Azure DevOps
- LangChain
Architecture
- Hybrid search and reranking over pgvector embeddings to improve relevance beyond naive similarity search.
- Ingestion and chunking pipelines for structured data (Excel / CSV), with end-to-end retrieval fully instrumented for observability.
- Responsible AI observability dashboard on Azure Databricks tracking model usage, input parameters and ethical-AI metrics.
- High-performance REST APIs with FastAPI; owned Azure DevOps CI/CD across all environments.
Result
A shared retrieval service with better relevance and full observability, and Responsible AI reporting to support transparency and regulatory compliance.
05 / Skills
Tools and technologies.
- Backend
- PythonFastAPIREST APIsPostgreSQL / pgvectorSystem design
- Frontend
- React
- Cloud & deployment
- AzureAWSDockerKubernetes (AKS)HelmAzure DevOps CI/CD
- Observability
- OpenTelemetryOpenInferenceArize AILocust
- Data
- Azure DatabricksMedallion architecture
- AI
- Azure & AWS model servicesAWS AgentCoreRAGLangChainLLM-as-Judge
Working style
How I work with AI tools
My strength is backend and platform engineering. For frontend work I pair with AI coding assistants such as Claude in VS Code — I own the design, data flow and review, and use the assistant to move faster on React UI. The AI Control Plane UI and this website were both built this way.