3+ yrsshipping AI inside Toshiba's industrial tooling
8+ LLMsbenchmarked on PLC code comprehension
0%stale evidence in EviCortex recall at 512 tokens (interim)
WinnerRapyd Global Hackathon · Intel Unnati finalist
The page is an agent
Don't read my CV. Run it.
Pick a question or type your own. It plans, calls tools over my real work and answers with sources.
scripted runs · instant
Answer
Choose a question above to watch a run.
01 — Orchestrate
Cascade AI
An open-source, three-tier agent CLI. An Administrator plans, Managers split the work, Workers execute, across Anthropic, OpenAI, Gemini, Azure and local models, with failover, permission escalation and per-tier budgets.
Now redesigning its orchestration around capability-aware routing, after a DeepSWE benchmark showed where a single model beat the hierarchy.
A local-first memory kernel for long-lived agents. It keeps every observation, retires superseded facts and compiles a cited context capsule under a token budget.
0%superseded sources at 512 tokens · baselines 100%
99.27%clean packs at 4k tokens · baselines 89.59%
Interim retrieval proxies on 200 knowledge-update queries. Paper in preparation.
Tests written in plain language for desktop, web and CLI apps. A multimodal model drives them, deterministic assertions decide pass or fail, and disposable VM capsules keep the host safe.
Built Enceladus: Mitsubishi GX Works3 programs in, AI-written documentation and a follow-up chat out. Owned PostgreSQL, FastAPI and React/TypeScript end to end.
Engineered a provider-agnostic agent system with a generate → respond → validate pipeline across Azure, OpenAI, Anthropic, Gemini, Copilot, Ollama and OpenAI-compatible endpoints.
Led an evaluation of 8+ LLMs on PLC code comprehension; drove an agentic workflow with OpenSpec and GitHub Copilot.
Shipped the engine that converts GX Works3 programs to IEC 61131-10 XML and Structured Text.
Jul 2023 – Jul 2024
Associate Software Engineer
Automated 700+ manual tests with Robot Framework and Python; Boofuzz fuzzing for robustness.
Delivered two PoCs on deadline, mentored 10+ peers and set up local Ollama LLM environments for the team.
Jan 2023 – Jul 2023
Trainee Engineer
Internal Python tools, a portable packaging tool and an automated Outlook reply system.
Finding · on-premise LLMs
Adding a third GPU made inference ten times slower.
2 PCs · 10GbE
40–45 t/s
+1 PC · 1GbE
3–5 t/s
Pooling office GPUs with llama.cpp RPC: activations hop from GPU to GPU in sequence, so the slowest network link sets the pace for the whole pool.