I design and build production-grade AI and data platforms, from LLM gateways, RAG systems, agents, and observability to lakehouse pipelines, cloud infrastructure, and cost-aware operations.

I work at the intersection of applied AI systems, large-scale data platforms, and cloud-native infrastructure.

• LLM platforms: multi-provider gateways, fallback routing, tool calling, agents, RAG, model serving, request telemetry, and token-level cost metering.
• AI reliability: agent reliability harnesses, grounded generation, RAG evaluation, LLM-as-judge workflows, regression gates in CI, MCP tooling, observability, and cost attribution.
• Data platforms: lakehouse architectures, batch and streaming pipelines, CDC ingestion, orchestration, schema-as-code, lineage, governance, and analytical workloads.
• Cloud platforms: infrastructure as code, containers, serverless workloads, CI/CD, observability, reliability engineering, and production operations across AWS, Azure, and GCP.

Reliability harness for tool-calling agents. It runs a real agent against frontier models and measures whether it stays grounded, refuses when it should, and avoids hallucinating, with cost and latency.

Highlights

• Production-shaped tool-calling agent (bounded loop, per-tool timeouts, citation-enforced grounding, explicit refusal and clarify)
• Tools served natively over MCP, the Model Context Protocol
• Live multi-model comparison across Claude Opus 4.8, Opus 4.7, Sonnet 4.6, and Haiku 4.5 on AWS Bedrock, with real cost and latency
• Outcome and trajectory metrics (grounding, hallucination, refusal, tool accuracy) gated in CI
• Surfaces where smaller models break, for example Haiku's hallucination rate jumps on ambiguous queries while the larger models hold

Stack: Python · Anthropic · AWS Bedrock · MCP · GitHub Actions

Production-style multi-provider LLM gateway designed for cost-aware and observable AI workloads.

Highlights

• Unified API for OpenAI, Gemini, Anthropic, and Amazon Bedrock
• Fallback routing across providers
• Token-level cost metering using integer micro-USD
• Append-only usage ledger for auditability
• Unified tool-calling interface
• Metered web research agent
• Prometheus metrics and Grafana dashboards
• Designed around production concerns: observability, cost attribution, and reliability

Stack: Python · FastAPI · OpenAI · Anthropic · Gemini · Amazon Bedrock · Prometheus · Grafana

CI-friendly RAG evaluation framework focused on retrieval quality, faithfulness, and regression control.

Highlights

• Retrieval evaluation metrics
• LLM-as-judge faithfulness checks
• CI regression gate that fails the build when quality drops
• Designed for repeatable RAG quality validation
• Useful for preventing silent degradation in AI applications

Stack: Python · BM25 · AWS Bedrock · GitHub Actions

• Building reliable LLM gateways for production AI workloads
• Designing RAG systems with evaluation, grounded generation, and quality gates
• Creating cost-aware AI infrastructure with telemetry, metering, and observability
• Applying platform engineering practices to AI and data systems
• Engineering lakehouse platforms with Spark, Iceberg, Delta Lake, Snowflake, and Databricks

LLM gateways · agents · tool calling · RAG · model serving · grounded generation · LLM observability · cost metering

Lakehouse architecture · medallion pipelines · CDC ingestion · orchestration · catalog & lineage · PII masking · multi-cloud migrations

Infrastructure as code · containers · serverless · event-driven systems · OIDC CI/CD · production operations

CI/CD quality gates · alarms & dashboards · request tracing · cost attribution · DLQ/quarantine flows · production runbooks

• Production-first design: systems should be observable, testable, deployable, and operable.
• Cost awareness by default: especially for LLM workloads, cloud compute, storage, and data movement.
• Reliability over demos: fallback paths, retries, dashboards, alerts, and runbooks matter.
• Evaluation-driven AI: retrieval quality, faithfulness, grounded generation, and regression gates should be part of the workflow.
• Data quality and governance: schemas, lineage, masking, validation, and ownership should be part of the platform.
• Automation and repeatability: infrastructure, pipelines, and deployments should be reproducible through code.