Senior software engineer + independent ML researcher
Building reliable systems and testing structured state for evidence-bound reasoning.
I work across production software, machine-learning evaluation, and independent research. My current research asks whether explicit bounded state improves evidence binding, memory, and transfer—and records where proposed mechanisms fail.
4 seeds
Frozen held-out transfer replicated
3,840
Finite evaluation traces validated
30 / 30
Frozen 7B decoder admission only
Current research
Structured state for conversational reasoning
A finite selector study produced exact four-seed transfer with a decisive state-disabled control. Follow-up studies then rejected the tested natural-language front end and bounded conversational architectures. A separately audited 7B decoder has now passed the language-substrate gate; the attachment comparison itself has not yet been run.
The positive result is finite, synthetic, and partly symbolic. It does not establish open-domain reasoning or a general theory of intelligence.
See the result, failures, and next experimentSelected engineering
Systems, applied ML, and automation
A curated set of projects chosen for technical scope and clear implementation evidence. Private prototypes and unfinished notes are intentionally excluded.
FlowTrot
A web/graph based low-code solution where you can create source code modules as stand-alone utilities or design more complex compute graphs.
Transformer-based Riptide Spotter
A hackathon project that uses a transformer model to detect rip currents from live beach camera footage. Mounted on a lifeguard tower, the system analyzes wave patterns to identify hazardous conditions in real time.
PadPilot
A macOS utility that turns a Bluetooth game controller into a mouse, keyboard shortcut surface, and local dictation controller.
Smart Home
A Home Assistant based smart-home control plane for local device automation, event handling, and natural-language assisted control experiments.
Research and engineering, together
My strongest contribution is the combination of long-running research ownership with production engineering discipline: explicit gates, reproducible artifacts, failure analysis, and systems that can be debugged end to end.