Applied AI / ML Systems
AI / ML engineer focused on agent systems, retrieval, evaluation, and production infrastructure. I build applied AI systems that make model behavior inspectable, measurable, and reliable in use.
Tool-using agent orchestration, retrieval reliability, and evaluation loops that hold up under real constraints.
AI & Agents: LangGraph, PyTorch, MCP, RAG
Architectures, Claude/OpenAI
Backend & Ops: Python, Go, Node.js, AWS,
Kubernetes, Docker
Data & Analytics: SQL, PostgreSQL, Vector DBs,
Airflow, Snowflake
01 / Work
Selected systems where orchestration, retrieval, evaluation, and deployment decisions shaped the result.
Production AI Systems
Finance-first A2A reasoning engine built on LangGraph and MCP.
Built a modular reasoning runtime for finance tasks where a single prompt loop was not enough. It separates intake, planning, context curation, execution, review, and output adaptation so evidence, assumptions, and policy checks stay explicit through the run.
System area Agent runtime, tool mediation, review flow, policy-aware orchestration
Reliable AI Systems
FHIR-grounded clinical summarization for the last 24 to 48 hours of patient change.
Built a healthcare RAG system on AWS Bedrock and LangChain for grounded summaries of recent patient change. The stack combined deterministic fact extraction, PHI redaction, and evaluation loops aimed at clinical reliability.
System area Retrieval, grounding, redaction, evaluation
Infrastructure / Sandboxing
Isolated execution environment for untrusted agent-generated code.
Designed a Kubernetes-based sandbox for untrusted agent-generated code with namespace isolation, resource controls, and a sidecar proxy for observability. The main challenge was safe execution boundaries, not simply running generated code in a container.
System area Isolation, throttling, execution safety, observability
Enterprise Retrieval / Automation
Schema-aware SQL automation for internal analyst workflows.
Replaced manual analyst requests with autonomous SQL agents backed by schema-aware retrieval and enterprise-safe execution. The core value came from constraints, RBAC alignment, and making the generated path trustworthy enough to use.
System area SQL generation, schema retrieval, policy-safe automation
Additional systems across web agents, efficient tuning, genomics, and streaming infrastructure.
Deterministic validation loop for web agents using vision-backed state checks.
RepositoryLightweight proxy tuning for LLaMA 2 to cut training cost without full-parameter finetuning.
RepositoryDistributed genomics inference and throughput optimization on HPC infrastructure.
Event-stream ingestion and agentic triage for disaster response mapping.
RepositoryReal-time meeting infrastructure with network recovery and higher peer capacity.
Repository02 / Notes
Notes are working records on failure patterns, implementation tradeoffs, and system behavior that held up in practice.
Systems note
Production failures in computer vision are rarely simple 'wrong predictions.' They are complex conceptual drifts—blur, lighting, camera shifts. Here’s how I built a 'flight recorder' to catch them before they become incidents.
03 / Experience
Roles most relevant to the systems work on this site.
Tabner Inc
Jul 2025 - Present
Built agent workflows for internal data operations, redesigned ETL pipelines, and shipped containerized services with production uptime requirements.
Virginia Tech
Aug 2024 - Jul 2025
Built genomics model workflows, distributed experiments, and repeatable fine-tuning pipelines across GPU infrastructure.
Virginia Tech
Jun 2023 - May 2024
Improved NER performance and inference efficiency while exploring lower-cost ways to steer foundation models toward domain tasks.
UJR Technologies
Jul 2021 - Dec 2022
Built backend APIs and delivery pipelines focused on reducing release defects and stabilizing integrations.
04 / Working Areas
Grouped by capability area rather than by tool lists.