utkarsh mauryasystems lab
01 / systems n = 06 pi 5 · rp2040 · risc-v bare metal · linux

systems

Six things I have built and measured, ordered by depth of engineering rather than recency. Every number here is tied to a commit and a capture; provisional results say so.

01
STATUSACTIVE
AREAOS / ROBOT.
PLATFORMARM64 / PI 5
TYPERESEARCH+ENG
YEAR2025–PRESENT
ORGVOLNLABS

AxiomOS

Measured

Robots freeze their kernels because one bad change bricks a fleet. AxiomOS makes kernel behavior a set of verified, hot-loadable programs, so a deployed machine can change what it does without reflashing and without leaving its safety envelope.

Verifier — minimal prog.
943.28ns
CRITERION 95% CI · 2ef74f0
Verifier — 1000 instr.
376.43µs
NEAR-LINEAR SCALING
Actuation monitor
37ns med.
PROVISIONAL · 5 µs BUDGET
VERIFICATION
In-kernel eBPF verifier using path-sensitive abstract interpretation, with an aarch64 JIT.
SCHEDULABILITY
WCET cost model over instruction classes feeding EDF utilization admission: attaching a program commits CPU time against a fixed budget, and the kernel refuses attachments that would exceed it.
ACTUATION
Reference monitor clamps every PWM and GPIO command to a per-profile envelope (duty ceiling, slew limit, e-stop latch) before it reaches MMIO. Zero-escape invariant established by unit tests and property-based testing.
EVIDENCE
A release-evidence gate ties every published number to a clean commit, a hashed artifact, and a committed capture. Results failing provenance are demoted to provisional and may not be cited — the gate has demoted prior headline results, including an earlier Linux comparison.
BRING-UP
GIC, ARM generic timer and RP1 GPIO/PWM/UART drivers written from scratch on Raspberry Pi 5; memory management, processes, ext2 VFS, syscalls. RP2040 co-controller as a hardware safety interlock.
Repository → Rust no_std · eBPF · aarch64
FIG. 01
Hardware-in-the-loop bench: Raspberry Pi 5 target, scope capture, RP2040 interlock under test.
Electronics test bench with oscilloscope, soldering iron and prototype board
02
STATUSMS. IN PREP.
AREAMEMORY MIGR.
PLATFORMCRIU 4.2 / UFFD
TYPERESEARCH
YEAR2026
KERNELS6.18.7 / 7.0.9

DistriProc

Reproduced ×2 kernels

Speculative paging in CRIU lazy restore. Whether sequential prefetch helps or hurts time-to-first-request is governed by network RTT — and fault count turns out not to be a proxy for latency.

TTFR — loopback, PyTorch
+88%
650 → 1227 ms · t = −45.9
Crossover RTT
~125µs
−37% TTFR AT 1 ms
Controller recovery
5.5ms
OF DEMAND-ONLY · p = 0.42
QUESTION
Under what conditions does prefetching pages ahead of demand actually shorten time-to-first-request during lazy restore?
FINDING
At loopback, prefetch increases PyTorch TTFR by 88% while cutting page faults 85%. The effect inverts near ~125 µs RTT; demand-only lazy restore times out entirely at 2 ms.
CONTRIBUTION
An adaptive userspace controller monitoring duplicate pressure and queue depth per fault window, disabling prefetch when it detects waste. Cuts prefetch volume 52–55% on memory-light workloads.
METHOD
20 iterations per configuration across three workloads, netem-shaped RTT sweep, Welch tests. Locked claims and an out-of-scope list committed alongside the artifact; two reproduction guides published.
Artifact → C · Python · userfaultfd · netem
03
STATUSUPSTREAM
AREALINUX KERNEL
ARCHRISC-V
TYPEENGINEERING
PATCHES4 · LANDING

Linux Multikernel

CPU lifecycle and spawn path for multikernel boot on RISC-V: bringing a second kernel up on parked harts without disturbing the first.

#23SBI HSM-based spawn, with confirm-parked and release
#24fence.i entry stub, avoiding a stale I-cache on re-spawn
#25spawn halt/reboot must HART_STOP every hart, never SRST
#26KEXEC_TYPE_MULTIKERNEL Image loader, placement in the grant, no purgatory
Tree → C · SBI · kexec
04
STATUSPATENT FILED
AREAEDGE / ROBOTICS
PLATFORMPI 5 / JETSON
TEAM7 PERSON
YEAR2025–PRESENT

ResQTerra

Flown

Search-and-rescue flights happen where network reliability cannot be assumed, so a drone that loses its link has to stay controllable and predictable rather than fail open.

Staged degradation on link loss — heartbeat timeout
10s
WARNING RAISED
30s
RETURN TO HOME
60s
EMERGENCY LANDING
FAILOVER
Companion computer holds a persistent session with ground control over 5G and falls back automatically to a Bluetooth relay node when the primary link drops.
PROTOCOL
Protobuf envelopes over length-prefix framing, shared by edge, server and relay as one crate. Mission start/abort, RTH, status and emergency motor kill, each acknowledged and timeout-tracked.
FLIGHT
MAVLink 2.0 bridge to ArduPilot over serial, UDP or TCP — arm/disarm, takeoff and land, guided waypoints, mission upload — streaming GPS, battery and flight-mode telemetry back. Patent filed on GPS-denied navigation for infrastructure-failed environments.
Engine → Rust · Tokio · prost · ROS
FIG. 02
Signal capture during driver bring-up.

FIG. 03
Breadboard interlock prototype, annotated pinout in frame.
Digital oscilloscope showing a sine and square wave capture on a lab bench
Desk setup with breadboard circuit, oscilloscope and printed pinout diagrams
05
STATUSMAINTAINED
AREAEDA / HW
TYPEOPEN SOURCE
STARS~55

Cirkitly Toolchain

An organization of AI-assisted hardware design and verification tools. Spec-driven RTL generation is only useful if the output is checked, so every tool in the set pairs generation with verification.

spec2rtl-pluginCLI copilot for spec-driven RTL generation and verification22★
spec2rtlArchived predecessor29★
code2testHierarchical test generation with human-in-the-loop verification—
picoplaceRust CLI for PCB layout synthesis—
rtl-copilotNatural-language Verilog with visual FSM editing—
Organization → Python · Rust · Verilog
06
STATUSACTIVE
AREARL / EVALUATION
ORGPOLYMATH
TYPEENGINEERING
YEAR2026–PRESENT

Agent Evaluation Systems

Production-style RL environments for training and evaluating frontier coding agents. The hard part is not the task — it is a verifier the agent cannot game.

ENVIRONMENTS
Build environments that mirror real systems work rather than toy repositories.
EVALUATION
Design tasks, sealed verifiers and failure-injection tests for complex systems problems.
REVIEW
Review tasks for realism, difficulty, reproducibility and reward-hacking resistance.
Polymath → Python · Docker · PostgreSQL · Redis