Systems Software Engineer · Mindfire Solutions → Razorpay → Amazon → Oracle → Microsoft
Rust · C++ · Go · CUDA · Kubernetes · eBPF
"Never settle, always regret, then refactor the mess."
"Age gele baaghe khai" (আগে গেলে বাঘে খায়): rush ahead, and the tiger eats you.
Full-stack hardware and software platform to build a distributed edge GPU cluster, curate data to fine-tune LLMs, and serve them through a multi-engine inference layer: vLLM, Ollama, TensorRT Edge LLM, llama.cpp, SGLang.
Data curation. 13,114 training examples (6.5M tokens, 35 sources), checked by
sparkagainst slop and the five Rust rules, with before/after scoring for the fine-tune.Secured hosting, bring your own domain. One encrypted key per person. A launch note states what is stored and what is not. voltforge.tech forwards to the cluster over Tailscale.
OpenAI-compatible model routing that every harness plugs into: Claude Code, openBatarangs, OpenCode, all on one API. The 30+ chapter book covers the cluster, the checker, the data pipeline, serving, security and keyring crypto.
⚡ GPU Parallel Programming from First Principles
16 chapters from the mathematics of parallelism to real kernels: Amdahl, Gustafson, the roofline model, warps, streaming multiprocessors, the memory hierarchy. CUDA C++ kernels and the same kernels in Rust with cuda-oxide. Reduction, scan, histogram, tiled matmul, an image-processing capstone, profiling with Nsight Compute. Every program run on a Jetson AGX Thor (sm_110) and Orin NX (sm_87).
📚 Cracking the Systems Programming Interview: books and labs in one repo
- Distributed Systems from the Ground Up: 28 chapters. Part 1 builds a replicated, sharded key-value store in Rust: write-ahead log, leader replication, Raft election, group commit, each measured against the failures it claims to survive. Part 2: 22 system designs with runnable simulations.
- Democratizing Linux: A Pipe Dream: 13 chapters that rebuild kernel mechanisms in user space, every program in Go and Rust: pipes, file descriptors, a shell, signals, CPU scheduling, virtual memory, epoll, TCP, namespaces, cgroups, a container.
- GPU-Accelerated Kubernetes: informers from scratch, a GPU job controller on controller-runtime and kube-rs, how GPUs are shared (time-slicing, MPS, MIG, vGPU, DRA), and a kind cluster whose pods use a real GPU. Go and Rust.
- The interview book: 19 chapters in six parts: tools, whiteboard algorithms, memory and the machine, concurrency, the network, interview day.
- Labs: 31 Rust modules and 22 programs (hand-rolled async runtime, epoll server, TCP and HTTP servers) with benchmarks and an 8-page report, a Kubebuilder
GpuWorkloadoperator, 14 study guides.
Agentic coding CLI for local models: auto-discovers the best Ollama / TensorRT / llama.cpp model on Jetson and edge hardware, then reads, edits, and verifies code with tools. No cloud, no manual model ops.
Edge-AI road safety, 100% on-device: YOLOv8/v11 → TensorRT (~12 ms) on Jetson Orin, kinematic decision engine, Kalman + Deep SORT, voice alerts, zero video leaves the car. vision model · pi-stream · stream-client · companion · data-crowd
🕸️ eBPF from the Ground Up · 📘 Modern C++ Memory-Safe Systems · 👁️ Seeing Machines · 🎮 Bevy Physics · ☸️ Kubernetes Operator Pattern · 🥊 Coding Interview FightClub
|
🦇 openBatarangs Rust · local agentic coding CLI |
🐯 thor-thunder-tigress Rust · edge GPU cluster · multi-engine serving |
⚡ gpu-parallel-book CUDA · Rust · 16-chapter book |
|
📚 cracking-systems-interview Books & labs · distributed · Linux · K8s |
🚗 civicsense-vision-model Rust · on-device AI perception |
🦀 cuda-oxide-demo Rust · CUDA-Oxide · Jetson benchmarks |
|
🕸️ ebpf-rust-networking-book Rust · eBPF/Cilium/K8s book |
📘 modern-cpp-systems-book C++ · memory-safe systems book |
👁️ seeing-machines-book Python → Rust · CV book |
|
🎮 bevy-physics-book Rust · physics in Bevy |
🔍 systems-patterns Go · systems hands-on |
🧠 AdvancedAlgoPatterns Kotlin · DSA patterns |
|
🥊 FightClub_600 Interview prep · 5 languages |
📡 jobsense-parker Rust · job-hunt CLI |
🎸 grungeGPT Rust · from-scratch GPT + BPE |
…plus 50+ more repos: kube-goAt, zero-trust-networking, andross, multicloud-distiller, distributed systems, ML notebooks, music & audio, websocket scaling, NFC + e-ink hardware, and a Neovim agentic IDE.
| Layer | Weapons of choice |
|---|---|
| Languages | Rust, C++, Go, Kotlin, Python, Java, TypeScript, C |
| Systems | Lock-free data structures, memory safety, kernel/eBPF, low-latency |
| Networking | TCP/IP, UDP, TLS/mTLS, HTTP/2/3, gRPC, QUIC, Cilium, XDP |
| Distributed | Kafka, Flink, Spark, consensus/replication/sharding, CDC |
| ML/AI | PyTorch, TensorFlow, BERT, ONNX Runtime, Ollama, edge inference, CUDA |
| Cloud | AWS, GCP, Azure, Kubernetes (EKS/AKS), Terraform, Kubebuilder |
| Company | Role | Impact |
|---|---|---|
| Microsoft | Software Developer II | Led legacy C++/C# → Rust migration: 40% lower P99, 2× throughput, 3× less memory, 6-figure savings. Built AI policy engine + zero-trust sovereign networking (eBPF/Cilium). |
| Oracle Cloud | Senior Software Engineer | Lock-free DB internals: +45,000 ops/s. Terraform provider for Autonomous DB (~$1.2M/mo savings). |
| Amazon | SDE II | Real-time ML ranking (XGBoost + DL): 100ms latency, 5% CTR lift, $5.4M/mo revenue. BERT inference @ 5,000 QPS. Petabyte-scale data platform. |
| Razorpay | SDE | Go UPI gateway: 1M+ hourly transactions, 99.99% uptime. |
LinkedIn · YouTube · GitHub · Portfolio · 📧 arpan.pathak47@gmail.com
Currently: teaching the Thor Tigress Cub on a Jetson AGX Thor to stop writing AI slop, shipping openBatarangs, and writing books that teach it all from first principles.



