Engineer building efficient ML systems and the infrastructure they run on
I'm interested in the intersection of model architecture and inference efficiency — why modern sequence models are designed the way they are, where they fail, and whether simpler alternatives can close the gap.
My main project is PRISM — a modality-agnostic sequence model with a 12-layer hybrid backbone interleaving S4D-Complex blocks (continuous-time signal dynamics) with Gated Delta Rule blocks (matrix-valued associative memory) at a 3:1 ratio. The same backbone handles 12-lead ECG, image patches, and arbitrary continuous signals — only the input projection and output head are per-modality. Hybrid configuration reaches 88.4% on CIFAR-10 with ~8M parameters. Block-pattern ablations and PTB-XL ECG benchmarks are in progress; results land in the README as runs complete.
Alongside the architecture work, I build the systems that production ML lives in — Go/Rust microservices, observability pipelines, and infrastructure tooling. I think the gap between "model that works in a notebook" and "model that ships" is a real engineering problem worth being good at.
Current focus areas:
- State-space models (S4D, complex diagonal SSMs) and associative memory (delta rule, fast weights)
- Hybrid sequence architectures — when interleaving beats either component alone
- Hardware-aware algorithm design (parallel scan, chunked recurrence, memory hierarchy)
- Modality-agnostic backbones — testing how far a single design generalizes across signal types
- Production ML infrastructure — agents, observability, distributed services
PyTorch SSM Delta Rule Research · Active
A modality-agnostic sequence model: 12-layer hybrid backbone interleaving S4D-Complex blocks with Gated Delta Rule blocks at a 3:1 ratio. Custom parallel scan implementation (work-efficient Blelloch upsweep/downsweep) for the SSM path; chunked recurrence for the delta rule. Same backbone handles ECG, image patches, and continuous signals — only the input projection and output head change per modality. Hybrid run: 88.4% on CIFAR-10 (~8M params). Ablations and ECG benchmarks pending.
PyTorch Research · Study repo
Earlier exploration of a single O(n) primitive (local convolution + linear attention + gated fusion + key-value memory) as a transformer alternative. Predates PRISM and reflects an earlier mental model — kept public as a reference for the design choices that led to PRISM.
Go Rust TypeScript Docker · Production-grade
A full-stack monitoring and control platform for distributed services: Go backend (auth, metrics, alerting, SLO tracking, incident management), Rust agents for low-overhead host-side data collection, React/TypeScript frontend, and a mobile client. ~70k lines across the stack. Built solo end-to-end as an exercise in shipping production systems — not just designing them.
| ML / Research |
|
| Systems |
|
| Web |
|
| Infra |
|
