Modern ecosystems are changing faster than humans can observe. Field researchers rely on slow, manual measurements, while satellites offer delayed, low-resolution snapshots. There’s a massive sensing gap between the ground and the sky.
My mission: build the robotics stack that makes ecological monitoring cheaper, longer, repairable, and more intelligent; drones, sensors, and edge-AI systems for real-time environmental understanding.
I design the entire stack myself:
C-based transformers → HPC training pipelines → embedded flight controllers → sensors → ecological applications.
Every prototype I publish — BLDC controllers, CPU-optimized LLMs, IMU calibration rigs, test benches; is a building block toward the world’s most accessible ecological intelligence system.
If you’re interested in:
Embedded AI on CPUs
Drones, flight dynamics, attitude control
C/MPI/HPC engineering
Real-world conservation through robotics
Deep technical builds from first principles
This channel is for you.
ANTSHIV ROBOTICS
Two PHP memory-safety findings, one practical question: what state is visible when allocation fails or a request ends?
Swipe through the half-published bucket, the request/worker lifetime boundary, the tiny C teaching model, and the upstream repair. The comic is an analogy, not a literal heap layout or exploit recipe.
Full article, code, primary sources, and diagrams: www.shivasnotes.com/blog/6003/inside-php-s-2026-me…
#PHP #MemorySafety #CProgramming #SecurityEngineering #ShivasNotes
1 week ago | [YT] | 0
View 0 replies
ANTSHIV ROBOTICS
Your model does not read words. It reads token IDs.
This 10-slide walkthrough follows text through WordPiece, ranked BPE, SentencePiece Unigram, and registered special tokens. The ranked-BPE counterexample is the surprise: a fast longest-vocabulary lookup can produce different IDs from the model's learned merge order.
Read the full illustrated math log and try the interactive merge exercise:
www.shivasnotes.com/blog/6000/from-wordpiece-to-ra…
CKE tokenizer guide and source: c-kernel-engine.github.io/C-Kernel-Engine/tokenize…
Join the ShivasNotes Discord from the site to discuss tokenizer parity and CKE.
#Tokenization #BPE #WordPiece #SentencePiece #CKE
2 weeks ago | [YT] | 2
View 0 replies
ANTSHIV ROBOTICS
A drone camera needs a 45° yaw correction. How do you turn its current and target quaternions into an axis and angle the controller can reason about?
This 10-slide walkthrough starts with the physical turn, shows why the quaternion stores a 22.5° half-angle, recovers the +Z axis and 45° angle, then checks zero- and 180-degree edge cases. The linked article includes animated diagrams, a compiled C example, and the implementation history in attitudeMathLibrary.
Full worked article:
www.shivasnotes.com/blog/5997/quaternion-to-axis-a…
Attitude math library and tests:
github.com/antshiv/attitudeMathLibrary
#Robotics #Quaternion #Drone #AttitudeControl #Math
2 weeks ago | [YT] | 3
View 0 replies
ANTSHIV ROBOTICS
Why is 4-bit quantization so effective, and why are Q4_K, MXFP4, and NVFP4 not interchangeable? This 10-slide walkthrough follows the bytes: a 64-byte cache line, scaled four-bit codes, two codes packed into each byte, reconstruction error, and the wider math inside the kernel.
The full worked article includes format layouts, CPU and GPU execution paths, CKE implementation links, and what to benchmark before claiming a win:
www.shivasnotes.com/blog/5996/why-4-bit-quantizati…
CKE quantization documentation:
c-kernel-engine.github.io/C-Kernel-Engine/quantiza…
Follow ShivasNotes for kernel math, model internals, and practical systems experiments. Join the CKE community through the links in the article.
#Quantization #LLM #CKE #MachineLearning
2 weeks ago | [YT] | 4
View 0 replies
ANTSHIV ROBOTICS
What does Muon actually change when training a model?
Start with one wx+b gradient. Batch examples into a matrix, follow the backward GEMM, and compare the first update from SGD, AdamW, and Muon using the same 2 x 2 gradient. Then look at optimizer state, Newton-Schulz, memory, and what a small training run can and cannot prove.
Full worked article, figures, code, and experiment caveats:
www.shivasnotes.com/blog/5995/muon-optimizer-from-…
Explore C-Kernel-Engine, the open-source C training and inference project:
github.com/C-Kernel-Engine/C-Kernel-Engine
CKE community and Discord onboarding: the contributor page explains how to request an invite and where to take technical discussions:
c-kernel-engine.github.io/C-Kernel-Engine/contribu…
DeepSeek-V4's optimizer report, discussed in the article:
arxiv.org/abs/2606.19348
The toy comparison is not a benchmark of CKE's Muon support; CKE does not currently ship the Muon kernel used in these plots. The article's NumPy examples use fixed teaching settings, not a tuned optimizer contest.
#Muon #Optimization #MachineLearning #KernelMath #CKE
3 weeks ago | [YT] | 2
View 0 replies
ANTSHIV ROBOTICS
A model can answer a prompt while an agent still fails to complete the task. Why? The tokenizer, publisher chat template, tool protocol, server response and client harness all have to preserve the same contract across turns.
This 10-slide carousel follows the C-Kernel-Engine serving path: an early contributor-led DS4 harness attempt, a measured Qwen Code tokenizer mismatch, tool admission, the harness-side file read, observation replay, and the difference between controlled HTTP tests and a real agent task. It does not claim the early DS4 attempt was a retained benchmark or that every model/harness pairing is certified.
Read the full article and linked CKE PR history: www.shivasnotes.com/blog/5994/how-agent-harness-wo…
CKE repository: github.com/C-Kernel-Engine/C-Kernel-Engine
Keywords: C-Kernel-Engine, CKE, agent harness, model serving, tool calling, tokenizer parity, chat templates, Qwen Code, Muse, Instella
#CKE #AgentHarness #ModelServing
3 weeks ago | [YT] | 1
View 0 replies
ANTSHIV ROBOTICS
How does a convolution filter learn? This 10-slide walkthrough starts with one five-pixel row, follows the loss gradient through shared weights and overlapping inputs, then shows a real NumPy gradient-descent run.
The hand-worked Conv1D example moves loss from 1 to 0.9322 after one SGD step. A separate synthetic supervised edge task uses all 32 binary five-pixel rows (96 windows) and moves loss from 0.25 to about 1.08e-14 over 300 full-batch steps. Its learned filter approaches [-1, 0, 1]. This demonstrates the gradient calculation and a deliberately defined local target, not real-image recognition or spontaneous feature discovery.
Read the full article, math, plots and runnable code:
www.shivasnotes.com/blog/5992/backpropagation-thro…
Part one, Convolution Is Wx+B Moving Across a Signal:
www.shivasnotes.com/blog/5991/convolution-is-wx-b-…
C-Kernel-Engine:
github.com/C-Kernel-Engine/C-Kernel-Engine
#Convolution #Backpropagation #KernelMath #CKE #CPUAi
3 weeks ago | [YT] | 3
View 0 replies
ANTSHIV ROBOTICS
Convolution is old signal-processing math, but it still helps modern AI models see and listen.
I continued my wx+b and GEMM series with a small weighted sum that moves across a signal. In ten slides: one RGB pixel versus its neighbourhood, the exact 1D arithmetic, why shared weights find the same edge in different places, a 2D filter, the GEMM view and its memory cost, and where these kernels sit in Whisper, Kokoro, Qwen3-TTS and vision encoders.
The full article has worked numbers, Matplotlib plots, Python examples, CKE C-kernel source links and the PRs that introduced the audio convolution paths:
www.shivasnotes.com/blog/5991/convolution-is-wx-b-…
CKE kernel documentation:
c-kernel-engine.github.io/C-Kernel-Engine/audio-ke…
CKE source:
github.com/C-Kernel-Engine/C-Kernel-Engine
Convolution is not the entire model, and not every vision encoder uses Conv2D. The point is to understand the local operator, its alternatives, and the contract connecting it to the rest of the graph. Part two will cover backpropagation through shared and overlapping windows.
#CKE #Convolution #KernelEngineering #ComputerVision #SpeechAI
3 weeks ago | [YT] | 4
View 0 replies
ANTSHIV ROBOTICS
Three audio models. Two CPU systems. The same 42:22 recording. No GPU.
C-Kernel-Engine now completes Whisper Base, NVIDIA Parakeet TDT 0.6B v3, and Cohere Transcribe 2B through generated C model circuits. I ran the same source audio at the same CKE revision on my Intel P3 and AMD Ryzen nodes.
Whisper Base
P3: 121.74 seconds, 20.88x real-time
Ryzen: 65.25 seconds, 38.96x real-time
Parakeet TDT 0.6B v3
P3: 1,515.19 seconds, 1.68x real-time
Ryzen: 881.95 seconds, 2.88x real-time
Cohere Transcribe 2B
P3: 1,280.99 seconds, 1.98x real-time
Ryzen: 583.02 seconds, 4.36x real-time
The transcript artifact for each model matched exactly between Intel AVX2 and AMD AVX-512. That demonstrates deterministic cross-host execution for these retained runs. It does not claim that the three models have equal transcription quality. Cohere also uses a pinned external speech schedule and infers 2,288.04 seconds of the recording, so its timing policy differs from the other paths.
Whisper is practical in my production workflow today. Parakeet and Cohere now work end to end through CKE's generated code, and the measurements expose where CPU utilization and kernel performance still need work.
Read the complete article:
www.shivasnotes.com/blog/5981/from-native-kernels-…
Watch the exact 42:22 source video used for the benchmark:
https://youtu.be/XMyPVVmjuiY?si=gEwVi...
Watch the latest follow-on video about the networking ideas that influenced CKE:
https://youtu.be/cHUGuZs4htw?si=KFxu1...
CKE hardware lab and the roles of the P3 and Ryzen nodes:
c-kernel-engine.github.io/C-Kernel-Engine/support-…
C-Kernel-Engine:
github.com/C-Kernel-Engine/C-Kernel-Engine
#CKE #CPUAi #Whisper #Parakeet #Cohere #SpeechRecognition #KernelEngineering
1 month ago | [YT] | 1
View 0 replies
ANTSHIV ROBOTICS
How do you get started with C-Kernel-Engine (CKE)?
This carousel follows the smallest useful path: prepare Linux, clone CKE and its submodules, run a small Gemma 3 model on a CPU, inspect the generated C pipeline, and understand what one coherent result does and does not prove.
The easiest prerequisite is an older x86-64 laptop or desktop you can repurpose as a native Linux node. This keeps the first experiment inexpensive, gives you direct access to compilers, perf, CPU affinity, NUMA and memory controls, and turns learning Linux into a useful side quest. It also lets you build and inspect AI software on compute you own instead of making every experiment depend on a rented platform.
CKE combines explicit model circuits, kernel capability maps, memory planning, generated C, native CPU execution, and numerical evidence. It is still an emerging research system, so the artifact, quantization, context length, CKE commit, processor, and tested boundary all matter.
Read the complete getting-started guide:
www.shivasnotes.com/blog/5979/how-to-get-started-w…
New to the architecture? Start with What Is the C-Kernel-Engine?:
www.shivasnotes.com/blog/5889/what-is-the-c-kernel…
C-Kernel-Engine repository:
github.com/C-Kernel-Engine/C-Kernel-Engine
Official CKE quickstart:
c-kernel-engine.github.io/C-Kernel-Engine/quicksta…
CKE concepts:
c-kernel-engine.github.io/C-Kernel-Engine/concepts…
CKE v8 runbook:
c-kernel-engine.github.io/C-Kernel-Engine/v8-runbo…
Model and kernel evidence matrix:
c-kernel-engine.github.io/C-Kernel-Engine/model-ke…
#CKE #CPUAi #KernelEngineering #OpenSourceAI #Linux
1 month ago | [YT] | 8
View 0 replies
Load more