Who's running the process
The part of me that a hiring manager meets first.
I'm Mokshagna — most people call me Moksha. I'm a systems engineer in Bengaluru, and I spend my time in the part of computing where the abstractions run out and you have to deal with the machine as it actually is.
Right now that means automotive platform work at Jaguar Land Rover — three operating systems sharing one ARM SoC, where a 600 millisecond regression in first-screen render is a real problem someone has to go find. Alongside it: kernel memory management, a compiler down to x86-64, a zero-copy IPC bridge, and a limit order book engine chasing sub-microsecond adds. The one above is that same design in JavaScript, so you can watch it work.
Embedded systems and trading infrastructure look like different worlds, and people are usually surprised I care about both. They're the same job. Both are hard-real-time. Both punish you for allocating on the hot path. Both come down to knowing exactly what the hardware is doing on every cache line — and being stubborn enough to keep looking after everyone else has decided it's fast enough.
I should be honest about the process: I start one thing and finish somewhere else. A boot investigation turns into reading lk-boot disassembly at midnight. The detours are usually where the good part was, so I've stopped fighting it and started writing them down.
What I ship, and who pays for it
Automotive platform and OS internals. Three operating systems, one SoC, a boot budget measured in milliseconds.
Multi-OS platform, ARM Exynos Auto
Jul 2025 — presentJaguar Land Rover TBSI · Graduate SW Engineer · QNX, Linux, AOSP, C++, SELinux, cgroups, Yocto
- Cut the boot KPI to 4.8s across concurrent QNX/Linux/Android stacks via devb-ram UFS piggybacking and asynchronous SLM dependency chaining.
- Moved storage init into a native C++ service (
persist_mount), usingfork/execvto runblkidandresize2fssafely through custom SELinux domains during block-level migration. - Extended AOSP
libprocessgroupso native cgroup (cpu,blkio) constraints propagate into the Java framework layer, isolating resources across the app lifecycle to prevent starvation. - Relocated core components into secondary IFS images and moved the primary IFS into high memory — 600ms off first-screen render, primary IFS down to 32 MB.
- Reverse-engineered
lk-bootto map UFS storage routing and break the single-core sequential loading constraint in early init. - Enforced SELinux domain transitions for LXC containers —
lxcpowercontext inheritance,restoreconrelabeling, patched Treble compliance tests. - GTest coverage for the Health Monitoring System at 92.9% line / 90.5% function; GDB core-dump analysis for segfault triage.
Dynamic partitioning for OTA
May — Jul 2024Jaguar Land Rover TBSI · Software Intern · AOSP, Device Mapper, C++, Bash
- Designed a dynamic read/write partitioning scheme on ARM Exynos Auto, bypassing fixed ECU constraints with a custom AOSP super-partition for OTA updates.
- Built a native C++ tool and Bash CLI to manipulate AAOS Device Mapper mount paths, reconfiguring partitions on the fly at zero runtime cost.
- Validated across 50+ automated expand/shrink cycles with zero data loss; co-authored the internal white paper that made it the baseline for future ECUs.
Things I built because I wanted to know
Source on github.com/Moksha2004. Each one started as a question I couldn't answer from reading.
Limit order book matching engine
v1 · measured2026 — ongoinglowlat-orderbook · C++20, CMake · single-symbol, price-time priority, zero heap allocation on the hot path
- v1 comes in under budget:
add_orderp50 —,cancel_orderp50 —, against a 200 ns target. Timed with rdtsc over 200k iterations; full percentiles and the environment are published on the engine panel above, straight from the benchmark's own output. - Zero per-order heap allocation on the hot path, and not as a claim — an
LD_PRELOADmalloc counter runs the proof: 100k add/cancel cycles, 0 allocations attributable to orders. - Flat array price ladder for mid ± 1000 ticks with a
std::mapfallback beyond it: O(1) indexing where the volume actually is, correctness without ever resizing on the hot path. - Intrusive doubly-linked FIFO per price level — orders live in a pre-allocated arena, so addresses stay stable for the
OrderID → Order*map that makes cancels O(1). - Self-trade prevention (CancelNewest) and modify semantics that only surrender time priority on price change or size increase, matching CME/NASDAQ behaviour.
Zero-copy shared-memory bridge
Jan — Mar 2026C++, Python, pybind11 · POSIX mmap, lock-free SPSC ring, AF_XDP, NumPy buffer protocol
- Zero-copy C++ ↔ Python transport over POSIX
mmap, skipping kernel context switches entirely — 1.18B msgs/sec throughput. - Lock-free SPSC ring buffer using acquire/release ordering, 64-byte cache-line padding (
alignas(64)) against false sharing, and_mm_pause()backoff. - AF_XDP kernel-bypass ingest of raw UDP straight from the NIC into memory-mapped UMEM, eliminating per-packet syscalls and the kernel stack.
- Raw C++ memory exposed through pybind11 and the Python buffer protocol for vectorized NumPy reads at 850M msgs/sec — 500× over OS pipes.
Kernel memory management and tracing
Aug — Nov 2023GemOS · x86-64 · IIT Kanpur, Prof. Debadatta Mishra
- Implemented
mmap,munmap,mprotectandcforkwith lazy allocation and copy-on-write, including VMA partitioning, coalescing and access-flag validation. - Wrote the page-fault handler against a 4-level page table (48-bit VA), mapping physical frames and refcounting PFNs for CoW sharing.
- Built a circular trace buffer backing Linux-style
straceandftrace, intercepting invalid-opcode faults to capture syscall arguments and reconstruct nested backtraces.
Python subset → x86-64 compiler
Jan — Apr 2024Flex, Bison, x86-64 · IIT Kanpur, Prof. Swarnendu Biswas
- End-to-end compiler emitting x86-64 assembly: Flex lexer, Bison parser, optimized three-address-code IR.
- Hierarchical symbol table plus a code generator handling register allocation and runtime stack frames for deep recursion.
- Semantic analyzer enforcing static type safety across multi-level inheritance, overloading, typecasting and control flow, with traced errors.
Decisions, and the detours
Notes are decision logs — two designs were defensible, here's why I picked one. Detours are the other half: I start one thing and finish somewhere else, and the somewhere else is usually where the interesting part was.
Where software meets my hands
Drawings, paint, and things with wires. None of it is on my résumé and all of it is why I'm any good at the rest.
Drawing is looking at something long enough to see the structure underneath, then removing everything that isn't it. That's also what a profiler is for.
Bedrock
The numbers that aren't self-reported, what I reach for, and how to reach me.
Open to
C++, C, Python, x86-64 and ARM assembly, Bash
Linux kernel, POSIX, QNX and RTOS, AOSP, virtual memory, IPC, lock-free concurrency, SELinux, cgroups
CMake, Yocto/BitBake, Docker, GTest, GDB, pybind11, Flex/Bison, Git