Anthony Shivakumar

Technical lead on AI-enabled public-service systems, focused on CPU inference, compute efficiency, accuracy validation, and practical applications that can run inside government infrastructure.

• Worked with the Information Management Branch from 2022 to 2025, then moved to Connected Service BC from 2025 to present.
• Built and deployed CPU inference capabilities for government infrastructure, focusing on practical service consumption rather than abstract demos.
• Produced technical validation work comparing hardware and software options for AI workloads, including compute, accuracy, operational fit, and procurement implications for Xeon 6 dual-socket infrastructure.
• Worked on OCR and document-processing workflows for public-service forms, including SDPR-oriented form-processing pipelines.
• Collaborate across technical and non-technical teams to ship maintainable public-service software where reliability, auditability, and operational clarity matter.

Designed and built a multi-tenant HMVC web platform from scratch, running in production for 10+ years across multiple live sites on AWS EC2. Architecture supports multi-domain, multi-site, and multi-module tenancy through a custom PHP/Phalcon framework.

• Started from SEO and marketing services, then expanded into reusable internal software for websites, content, forms, sales, and operations.
• Built tools for generating and managing websites from structured content and reusable templates.
• Designed workflows for small businesses and independent creators who need practical web systems without heavy enterprise software.
• Run the platform as a real working product while continuing to improve the underlying architecture.
• Engineered a 3-tier HMVC architecture with MySQL, MongoDB, Docker, and Nginx — handling multi-tenant routing, module isolation, and per-site configuration at the platform level.

Worked on telecom router software for Ericsson SE edge routers and SSR smart services routers, supporting scalable fixed and mobile network infrastructure.

• Specialized in access control lists, packet classification, classification daemon workflows, and forwarding-plane behavior.
• Worked on software that interpreted CLI configuration, built lookup tables, and passed classification messages into routing-plane services.
• Implemented and maintained lookup behavior across software and hardware lookup paths, including algorithms such as linear lookup and MRFC.
• Worked around packet classification actions such as filtering, QoS class application, forwarding, NAT, RPF behavior, and carrier-grade packet-processing concerns.
• Used C, Python, assembly programming, shell scripting, Git, Gerrit, and Linux user-space and kernel-adjacent development practices.

Worked in C on access control list (ACL) and packet-classification software for Ericsson router platforms, with a focus on classification daemon behavior, troubleshooting visibility, and Python-based unit tests.

• Worked on ACL classification behavior in the router plane, including how classification configuration was interpreted and applied.
• Designed and improved classification and operator logging to help operators and developers triage ACL and packet-classification issues.
• Supported troubleshooting workflows for classification daemon behavior, configuration problems, and router software diagnostics.
• Worked with Python unit tests and related test tooling around classification and ACL behavior.

Personal technical site for writing about software, systems, AI, math, robotics, and the process of building practical tools.

• Used as a public home for technical essays, build logs, notes, experiments, and project documentation.
• Built toward a Simon Willison-style personal site: writing, tools, references, technical experiments, and useful public artifacts.
• Connects personal publishing with the same platform work behind ANTSAND.

From-scratch C systems project for CPU-native transformer training and inference, focused on hardware-aware AI, numerical validation, SIMD acceleration, quantization, observability, and distributed CPU scaling.

• Built a C-based transformer training and inference engine with hardware-aware execution on Xeon-class CPUs, reaching reported 474 GFLOPS benchmark results.
• Implemented and validated performance-oriented paths including AVX-512/AMX SIMD, NUMA-aware memory management, custom BPE tokenization, and Q4K/Q6K-style quantization experiments.
• Validated numerical behavior against PyTorch-style references with tight parity checks in the 1e-7 to 1e-12 range and reproducible benchmark reporting.
• Documents a roadmap for multi-version CPU-native model training and distributed execution, including larger-model experiments targeting 10B-100B parameter scales.
• Uses ShivasNotes as the public build log for documenting architecture decisions, benchmark results, failures, and roadmap work.

Embedded systems, robotics, drone, and physical-computing exploration connecting software, electronics, automation, controls, and practical experimentation.

• Explores robotics and drone projects through hardware-connected experiments, embedded software, sensors, controls, and automation workflows.
• Documents experiments through build notes and YouTube-style technical videos to make the learning process visible.
• Provides a practical base for robotics, electronics, drone, and embedded-systems consultation or prototyping work.