I work on a narrow question: how do you verify what a language model produces, rather than just trust it?
That question runs through my undergraduate thesis, my reading, and NOERΛVEN, the technology organization I'm building on the side. I build close to where software actually runs: Python for evaluation tooling, C, C++ and Rust for systems, shell for everything in between, with deep Linux fluency underneath all of it.
Praxis Decision: Sparse Upcycling and Committee Distillation of Julia-1 for Typed Decisions. Submitted to arXiv [cs.AI], October 2026.
A sparse mixture-of-experts adaptation of the Julia-1 typed decision model, fine-tuned on document relations, event relations, and evidence-conditioned claim verification. Distilling a probability committee into the model reached 76.5% on a 600-question holdout, 3.3 points above a matched hard-label control.
An early-stage technology organization spanning AI systems, firmware, and tooling, designed with safety as a first principle rather than an afterthought.
The developer and administrative console for NOERΛVEN services: one place to manage accounts, projects, and infrastructure.
Chat with models from multiple providers, organize conversations into projects, and bring files and sources into the discussion, so the next step stays close to the context that shaped it.
How software moves complexity upward.
Scarcity, labor, and value after artificial general intelligence.
Brought hardware configuration discrepancies close to zero with automated tracking across production departments. Processed 60–90 laptops and peripherals a day for cross-functional teams and resolved 20+ hardware, network, and system tickets across Windows and ChromeOS.
Maintained the RFID server behind tap processing for 1,000+ student records at Saint Scholastica Academy–Tabunok and Don Bosco. Ran daily MSSQL backups and fixed the SQL in the data-entry and ID-printing pipeline, ending data corruption during collection.