AI & agentic systems
LLM applications, retrieval, tool use, and memory—with explicit boundaries around what a model may decide.
Engineering & execution
Good ideas deserve
working systems.
I’m a software engineer working across Python, Java, backend systems, and AI applications. I care about the whole path: a clear problem, sound architecture, and a useful experience.
A practical way to build.
Make the problem specific. Keep the system’s decisions inspectable. Handle failure deliberately. And make the result clear to the person who needs to use it.
Capabilities
Depth where it matters.
Range to connect it.
My background includes professional Python and Java work, APIs and databases, and LLM applications. These are the areas I bring together in my projects.
Backend & data
APIs, structured data, durable workflows, and the architecture that turns an experiment into a dependable system.
Product engineering
The full path from a useful problem to a clear interface, with attention to usability, failure states, and evidence.
Applied experimentation
Exploring models and data through notebooks, evaluation, and small projects that make assumptions testable.
Selected case studies
Engineering, in context.
01 / Public project
HARK
Give the next run something better than a chat history.
The problem
An agent can repeat a procedure without retaining what worked, what failed, or why a recovery helped in a particular environment.
The engineering
A Python service stores run events and derived experience in CockroachDB. Scoped retrieval, failure fingerprints, provenance, and invalidation keep memory tied to the execution that produced it.
Scope & evidence
The public implementation focuses on CockroachDB query-regression diagnosis. Service and integration test sources are present; this portfolio does not claim independent live-runtime verification.
Inspect the source02 / Public project
Latch
Make the release decision inspectable.
The problem
Before a dataset crosses into an external AI system, ownership, eligibility, transformation choices, and approvals need to refer to the same evidence.
The engineering
The implementation separates bounded AI interpretation from deterministic checks and human approval. It includes a FastAPI service, a Next.js interface, DataHub context integration, and BigQuery measurement paths.
Scope & evidence
The focused workflow reviews support conversations for external model adaptation. Repository code and tests establish the implementation scope; a technical pass is not a certification of legal compliance.
Inspect the source03 / Public project
GridForge
Collection is the beginning. Compatibility is the problem.
The problem
A set of available components is not automatically a viable solar system. Voltage, current, storage, peak load, and budget must agree.
The engineering
A deterministic TypeScript solver enumerates bounded combinations, records constraint failures, and produces topology data for a single-line diagram. Scraper adapters keep inventory and source evidence separate from electrical validation.
Scope & evidence
The repository contains the solar constraint engine, solver tests, and a Next.js interface. A public demo may be linked after its landing page is checked; that check does not validate an entire live design run.
Inspect the sourceCollaborations & opportunities