I'm building Propeller at Abstract Machines — a WebAssembly orchestrator for deploying workloads across the cloud-edge continuum, from bare-metal servers to Zephyr RTOS microcontrollers.
The system is designed around a simple idea: edge infrastructure should be programmable with the same reliability as cloud infrastructure. The Go-based manager handles orchestration and scheduling, while lightweight Rust “proplets” execute workloads through Wasmtime, pulling OCI images over MQTT and running them close to where computation actually matters.
The most interesting work right now is around trusted execution. I'm building the TEE runtime layer so encrypted WASM workloads can be decrypted and executed inside Trusted Execution Environments, allowing clients to verify that their code ran exactly as intended and that sensitive data never left a protected boundary.
We also have federated learning built into the platform — training models across distributed edge nodes without moving raw data off-device. For me, this is where infrastructure becomes meaningful: not just distributed systems for their own sake, but systems that make secure, low-cost intelligence deployable in places where cloud assumptions fail.
I've been going deep on the WASM component model and how it reshapes the trust boundary between host runtimes and guest modules — especially relevant for confidential execution and TEE design.
I'm also studying the Confidential Containers attestation stack: how the Attestation Agent, Key Broker Service, and enclave verification pipeline work together so secrets are only released to verified runtimes.
On the Rust side, I've been learning to write async MQTT clients that stay under 1 MB resident on ARM Cortex-M4 targets. Constraints like that force a kind of engineering clarity you never get when building for cloud VMs with infinite memory and forgiving abstractions.
I keep returning to one question:
My first-author research modified YOLOv5s using Transformer encoder blocks and BiFPN to improve small pest bird detection in agricultural imagery, achieving a 4.8% mAP improvement with only 4ms additional inference time.
But model accuracy is only half the problem. Deployment is the real bottleneck.
With Propeller's WASI-NN integration, a WASM-packaged model can be scheduled to constrained edge nodes the same way as any other workload, with broad ARM compatibility and no dependency on constant cloud access.
The use case I care about most is African smallholder agriculture: early pest detection, drought stress monitoring, and low-cost field intelligence that works offline — using cheap SBCs, local inference, and infrastructure that assumes unreliable connectivity instead of pretending it doesn't exist.
That feels like the right problem: not building impressive demos for labs, but building durable systems people can actually use.