Arpan Pathak

Systems Software Engineer Β· I build fast things that don't break

🐍 Pythonβ˜• JavaπŸ“± KotlinπŸ¦€ Rustβš™οΈ C++πŸš€ Go⚑ CUDA🐧 Linux☁️ AWS☸️ Kubernetes
View Projects Download RΓ©sumΓ© Contact Me
πŸ“ Seattle, WA πŸ’Ό Open to work Β· H1B transfer 🐧 Linux Β· Edge AI Β· Local-First

01. About Me

Senior Software Engineer with 8 years building AI inference serving, backend microservices, and cloud-native infrastructure at Amazon, Microsoft, Oracle, and Razorpay.

Strong in Python, modern memory-safe C++, Rust, and CUDA: with BERT inference at 5,000 QPS (ONNX, INT8) on Docker/Kubernetes with CI/CD.

Passionate about computer vision & deep learning for road safety on NVIDIA Jetson Orin (TensorRT), lock-free data structures, eBPF/Cilium networking, and writing books about how it all works.

"Never settle, always regret, then refactor the mess." >আগে গেলে ΰ¦¬ΰ¦Ύΰ¦˜ΰ§‡ ΰ¦–ΰ¦Ύΰ¦―ΰ¦Ό
GitHub β†— LinkedIn β†— YouTube β†— Email β†—
0Years Experience
0BERT QPS Served
0Public Repos
0GitHub Followers
0% Uptime @ Razorpay
0Books Written

02. Ambitious Road Safety Project: CivicSense

🚦 FLAGSHIP · EDGE AI · 100% ON-DEVICE · PRIVACY-FIRST

Driving CivicSense: Edge AI Perception for Road Safety

CivicSense is an aftermarket edge-AI co-pilot for your windshield: it watches the road, understands traffic behavior, and talks to you, politely but firmly when civility slips. It detects intersection blockages, lane camping, stop-sign violations, road hazards, animals and debris, and broadcasts hazard beacons to nearby vehicles.

Everything runs 100% on-device: YOLOv8/v11 INT8 β†’ TensorRT at ~12 ms on an NVIDIA Jetson Orin Nano Super NX, a formal zero-training kinematic decision engine (dilemma zone, lead-vehicle, cut-in, stale rules), Kalman + Deep SORT tracking, and a voice-guided alert system. Built on 50,000 miles of real driving footage + LiDAR, with synthetic data from CARLA/SUMO.

Includes a published paper on Intersection Blockage Prediction, an ISO 26262 functional-safety study, Mobileye RSS-style safety reasoning, and an edge VLM serving stack (vLLM + TensorRT-LLM in Docker on the Jetson).

πŸ“„ Read the Paper β†— ⭐ Main Repo β†—

Git Submodules: The Full Pipeline

πŸ“Ή pi_stream β†’ civicsense-pi-streamThe capture layer. A Rust MJPEG + UDP live dash-cam server that turns a $15 Raspberry Pi Zero 2 W + Arducam into a phone-streamable camera. Hotspot mode, no cloud, no router, ~50 MB RAM. 🧠 pistream_client β†’ civicsense-stream-clientThe perception layer. A pure-Rust client (HuggingFace Candle, no Python, no ONNX Runtime) that pulls the Pi's MJPEG stream, runs YOLOv8n on every frame, draws boxes, and saves annotated frames. πŸ“± frontend β†’ civicsense-companionThe human layer. A Kotlin Multiplatform app (Jetpack Compose + SwiftUI) that subscribes to the pipeline's gRPC alert stream: intersection violations, lane warnings, stop signs and hazards, live on your phone. πŸ—‚οΈ datasets β†’ driving-civic-sense-data-crowdThe honesty layer. MIT-licensed training data + field-validation ground truth: 7-class YOLO labels, a formal ground-truth schema, validators, and public-dataset aggregation (COCO, BDD100K, UA-DETRAC, CARLA). No pixels in git.

03. Now Building: openBatarangs

πŸ¦‡ NEW FLAGSHIP Β· AGENTIC CODING Β· 100% LOCAL Β· JETSON-READY

Agentic coding CLI for local models β€” no cloud, no manual model ops.

openBatarangs talks to Ollama, auto-discovers the best model for your hardware, and iterates with tools to read, edit, and verify code. Built for Jetson-class edge devices and any laptop that runs Ollama.

πŸ¦€ RustπŸ€– Ollama⚑ JetsonπŸ”’ Local-FirstπŸ› οΈ Agentic Tools
⭐ Star on GitHub β†— πŸ“– Quickstart β†—

β–Ά Live Demo

2-minute screencast: setup β†’ auto model discovery β†’ real agentic coding session.

04. Books I've Written

05. Experience

Software Developer II @ Microsoft

Jun 2025 – Aug 2026 Β· Redmond, WA
  • Built zero-trust sovereign-cloud networking with eBPF + Cilium enforcing data residency & SecNumCloud compliance; wrote Kubebuilder operators reconciling state across AKS clusters in France & Germany.
  • Built & maintained Data Loss Prevention (DLP) and data lifecycle management services with a C#/.NET backend.
  • Refactored a legacy C++ codebase with modern C++ (smart pointers, RAII), fixing use-after-free and buffer-overflow vulnerabilities.
C++C#/.NETPythonKubernetesCiliumeBPFDockerLinux

Senior Software Engineer @ Oracle Cloud Infrastructure

Oct 2024 – Mar 2025 Β· Seattle, WA
  • Owned the Terraform provider (developer SDK) for Autonomous Database, replacing legacy control-plane tooling with a modern, maintainable SDK.
  • Built lock-free data structures that cut contention & latency in high-concurrency read-write paths.
  • Integrated OCI Security Vault with automated key rotation across control-plane services.
JavaGoREST/gRPCTerraformOCI

Software Development Engineer II @ Amazon

Mar 2021 – Oct 2024 Β· Seattle, WA & Hyderabad, IN
  • Scaled low-latency BERT inference to 5,000 QPS (ONNX Runtime, INT8), cutting triage SLA from 7 days to 2 hours.
  • Owned end-to-end training/serving for an XGBoost + DL ensemble across FreeVee, Twitch, MiniTV & Prime Video: +5% ad CTR, $5.4M/month revenue.
  • Led a team of 3 building programmatic-guaranteed ad delivery at <50 ms using real-time CDC (Kotlin, REST/gRPC).
  • Designed a uniqueness-constraint indexing SDK with two-phase commit & serializable isolation (3K writes/s).
  • Built a 1-petabyte warehouse with AWS Glue & Spark; cut deal over-delivery ~30% and overspend 25%.
PythonKotlinPyTorchXGBoostONNX RuntimeSparkKafkaDynamoDBRedshiftAWS

SDE @ Razorpay

Aug 2020 – Feb 2021 Β· Bengaluru, IN
  • Built the P2P Unified Payments Gateway in Go (Protobuf/gRPC): 3M+ transactions/hour at peak, 99.99% uptime, 150 ms P99 for Zomato, Swiggy, Zerodha, Groww.
GogRPCProtobufAWSGCP

Software Engineer @ Mindfire Solutions

Aug 2018 – Feb 2020 Β· Bhubaneswar, IN
  • Built Java/Spring full-stack tools and an AR product-search microservice handling 10K+ daily queries.
JavaSpringMicroservices

06. Featured Projects

πŸ§ͺ More Personal Projects

+ more on GitHub β†—

07. Skills & Toolbox

Languages

PythonKotlinJavaC++RustGoCC#/.NETTypeScriptJavaScriptShell/BashSQL

AI / ML Inference

PyTorchTensorFlowTensorRT (INT8)ONNX RuntimevLLMSGLangTensorRT-LLMBERTXGBoostYOLOv8/v11LLMsComputer VisionDeep LearningQuantizationMLOpsEdge Inference

GPU & Parallel Computing

CUDACUDA-Oxide (Rust→CUDA)SIMTGPU Kernel ProgrammingNVIDIA Jetson OrinHPCParallel Computing

Systems & Performance

Lock-Free Data StructuresConcurrencyMemory SafetyRAIISmart PointersLow-Latency SystemsComputer ArchitecturePerformance EngineeringMicroservices

Networking & Distributed Systems

TCP/IPUDPmTLSHTTP/2/3QUICeBPFXDPCiliumLoad BalancingKafkaSparkCDCgRPCRESTProtobufConsensusReplicationSharding

Cloud: AWS

AWS (EC2, ECS, EKS)LambdaKinesisEMRSageMakerCDKGlueCloudWatchDynamoDBRedshiftS3IAM

Cloud: Azure / GCP / Kubernetes

Azure AKSGCP GKEKubernetesDockerHelmTerraformKubebuilderIstioCI/CDPrometheusGrafana

Databases & Data

DynamoDBPostgreSQLMySQLRedisAmazon RedshiftAmazon S3SQLiteData WarehousingETL

08. Education

πŸŽ“

B.Tech, Computer Science & Engineering

RCC Institute of Information Technology, Kolkata Β· 2018

Data Structures & AlgorithmsOperating SystemsComputer NetworksDBMS

09. Let's Build Something

Currently engineering systems that make AI fast, safe, and affordable β€” building openBatarangs for local-first agentic coding β€” and writing books about how it all works.

βœ‰οΈ arpan.pathak47@gmail.com GitHub LinkedIn YouTube

πŸ“š Books I'm Writing

GPU Parallel Programming β†— eBPF + Rust Networking β†— CivicSense Vision Paper β†—