Software Engineer working on low-level systems, high-performance infrastructure, and things that have to be fast and correct at the same time.

I build systems from the ground up — a 16-bit CPU with a working pipeline and cache, an electronic exchange that replays real NASDAQ market data through an FPGA-style matching engine, Rust edge services handling 10M+ requests daily, and event-driven fintech pipelines. Currently contributing to NVIDIA's infra-controller (NICo), an open-source bare-metal/DPU lifecycle management system.

• ⚙️ Computer architecture — ISA design, pipeline hazards, cache modeling, FPGA-style dataflow
• 🦀 High-performance systems in Rust (Tokio, Actix Web, lock-free data structures)
• 📈 Market data and exchange internals — binary protocols, order books, latency measurement
• 📡 Event-driven architecture with Apache Kafka at production scale
• ☁️ Cloud-native infrastructure — Docker, Kubernetes, AWS

A large-scale Rust systems codebase for bare-metal and DPU lifecycle management. Contributing by navigating unfamiliar production code, root-causing issues with real evidence rather than assumptions, and working directly with the maintainer team.

• Diagnosed and closed a test-coverage gap in the mock BMC test harness — one hardware type had no way to be tested as a host machine at all. Added the missing test infrastructure and a regression test. (Merged)
• Found and filed a MachineId-collision bug affecting four hardware mock types, backed by reproducible test evidence; confirmed and fixed by a project maintainer.
• Traced a reported regression to its exact root cause by reproducing the failure against a real local cloud-init installation (not just reading documentation), and proposed a fix currently under maintainer review.

Stack: Rust · Tokio · PostgreSQL · Kubernetes · gRPC

A C++20 exchange that replays a full day of real NASDAQ TotalView-ITCH 5.0 market data — 8.7 GB, 282 million messages — and reconstructs the order book message by message.

• Complete ITCH 5.0 decoder covering all 21 message types with per-type length validation; zero malformed messages across 282,229,684 parsed
• Order-reference resolution layer — five of the eight book-affecting ITCH messages carry only an order ID, so the receiver has to remember every resting order. Fixed-capacity open-addressing table, zero allocation during replay
• Matching engine modeled as a 5-stage FPGA pipeline: price-indexed ladder (slot = (price − base) / tick) instead of a searched table, best bid/ask via a priority encode over an occupancy bitmap, no heap on the hot path
• 2× faster than the STL control implementation with a tighter tail — 40 ns vs 82 ns mean, 55 ns vs 162 ns at p99. Batch-mean methodology, since the platform clock only resolves 100 ns
• Reproducible FNV-1a digests over every book operation and the full ladder — the same input file yields identical hashes on any machine, compiler or optimization level
• Reconstructed AAPL's book to a 4-cent spread at the closing bell, and independently recovered the after-hours earnings print

Stack: C++20 · NASDAQ ITCH 5.0 · Lock-free / allocation-free design · CMake · Ninja

A complete CPU built from first principles in Rust. Every design decision — instruction encoding, memory layout, interrupt handling — made deliberately and documented.

• Custom 35-instruction ISA with 16-bit fixed-width encoding and 5-bit FLAGS register
• Two-pass assembler with forward-reference resolution compiling .asm → binary
• 5-stage in-order pipeline (IF → ID → EX → MEM → WB) with RAW data hazard stall detection, flag hazard detection for conditional branches, and 2-cycle branch flush
• Direct-mapped write-through L1 cache classifying misses as cold vs conflict; demonstrates cache thrashing on bubble sort's non-sequential access pattern
• 6 assembly programs: bubble sort, binary search, Sieve of Eratosthenes, RPN stack calculator, Fibonacci, Factorial
• 49 integration tests · 5 versioned releases · CI: fmt + clippy + test on every push

Stack: Rust · Custom ISA · Pipeline Simulation · Cache Modeling · Systems Programming

Rust-based AI agent implementing the full plan → act → observe execution loop.

• Modular tool system with dynamic tool registration and extensibility
• Async architecture with Tokio for scalable multi-step task execution
• Memory-safe by design using Rust ownership model
• Published as a reusable library on crates.io
• ⭐ 12 stars — organic traction from the Rust community

Stack: Rust · Tokio · LLM APIs · Async

Languages Rust • C++ • Python • Java • Go • TypeScript • Solidity

Systems Pipeline design • Cache modeling • Binary protocols • Memory-mapped I/O • Allocation-free hot paths • NUMA / core pinning • Latency measurement

Backend Tokio • Actix Web • Spring Boot • FastAPI • JAX-RS • Hibernate

Infrastructure Kafka • Redis • Docker • Kubernetes • Terraform • AWS • GitHub Actions

Databases PostgreSQL • MySQL • Amazon S3 • Redshift • Snowflake

• 💼 LinkedIn: linkedin.com/in/raj-mandaliya-78a622249
• 🌐 Portfolio: rajmandaliya-portfolio.vercel.app
• ✉️ Email: [email protected]

💡 Open to Systems Software / SDE III roles in systems and backend infrastructure.