• Building a distributed marketing automation platform that publishes one video across fleets of Instagram and TikTok accounts: a cloud control plane (FastAPI, PostgreSQL, React) plans the work, and an on-prem runner drives real Android phones in parallel to execute it.
• Engineered the orchestration layer as a pure state machine: randomized per-account scheduling, tiered task prioritization, daily rate caps, longest-waiting-first fairness, and a work-claim protocol with heartbeats and automatic retries between the cloud planner and the runner.
• Shipped an internal video learning platform (Next.js, TypeScript, PostgreSQL) with semantic search over tutorials powered by vector embeddings, plus auth and S3 media storage, end to end.
• Built 10 TypeScript apps for OpenSwarm's in-app collection, including a multi-agent generator that learns your writing voice and drafts social content that sounds like you, a real-estate research copilot that turns an address into a plain-English property dossier, a pre-publish compliance linter, and a recall radar that matches live FDA, CPSC, and NHTSA feeds against the products you own — with more on my GitHub.
• Redesigned and rebuilt OpenSwarm's public landing page, owning the UI/UX overhaul of the company's web presence.

TechPython, FastAPI, PostgreSQL, React, TypeScript, Next.js, Docker, Cloud Infrastructure

• Building Optivia with two co-founders: a platform that breaks a software project into a dependency-aware task graph, builds a specialized AI agent for each task with only the context it needs, and feeds every verified result back into a shared context, so a team and its agents work from one source of truth and get sharper over time.
• Architecting the orchestration engine (Python, Claude Agent SDK): LLM-driven task decomposition with plan validation and repair, per-task context scoping, concurrent execution across the graph, and automated test verification with self-correction when a task fails.
• Designing the engine to govern cost as well as quality, routing each task across model tiers with per-task token budgets, and measuring decomposition quality through a dedicated eval harness.
• Building the desktop client (Electron, React, TypeScript): an interactive execution-graph interface that visualizes task dependencies, agent status, and build progress in real time, backed by local SQLite persistence.
• Shipping the MVP as a desktop app for solo builders and small teams, with the architecture designed from day one to scale into an enterprise platform where whole organizations share centralized context instead of each engineer running isolated agents.

TechPython, Claude Agent SDK, Electron, React, TypeScript, SQLite

• Joined UC Santa Cruz's AIEA Lab as a high school junior, completing a research onboarding in neuro-symbolic AI: symbolic reasoning, logic programming, retrieval-augmented generation, and techniques for strengthening logical inference in large language models.
• Built LLM-to-Prolog reasoning pipelines (Python, OpenAI API, SWI-Prolog) that translate natural-language problems into formal logic, reimplemented the Logic-LM paper, and wrote a backward-chaining inference engine with unification and resolution from scratch.
• Worked on the ProSLM language model platform alongside undergraduate and graduate researchers, helping unify multiple LLM research components into a single deployable system on Google Cloud Platform.

TechPython, Prolog, OpenAI API, Google Cloud Platform