Perception
IN PROGRESS
PHYSICAL SYSTEM
- Sensor interfaces
- Camera and inertial test inputs
- Timing and calibration
SOFTWARE SYSTEM
- Vision pipeline
- Detection and tracking
- Uncertainty handling
Make intelligent systems prove themselves on real hardware.
If your model is too opaque, too slow, numerically wrong, or trapped in a stack you do not control, we trace it from equations to kernels and measured execution.
Bring us the model, runtime failure, or control problem. We find the first binding constraint, build the smallest defensible fix, and leave you with code and evidence your team can inspect.
OWN THE HARDWARE · UNDERSTAND THE EXECUTION
Own the stack
Keep your hardware, data, and execution path under your control.
Find the first failure
Localize the first wrong tensor, kernel, memory path, or numerical assumption.
Make CPU practical
Turn model mathematics into generated C and measurable work on real processors.
Leave with evidence
Carry forward tests, profiling results, limitations, and reproducible commands.
Own the stack from model inputs to measured hardware
Antshiv connects model and system design to explicit circuits, native execution, and evidence that can be inspected rather than merely asserted.
ONE ENGINEERING SYSTEM
Three programmes, one engineering method
Equation → model → numerical contract → kernel or controller → native execution → hardware behaviour → measurement.
C-Kernel-Engine
A C-first compiler and runtime that turns model weights and circuit templates into inspectable generated C and native CPU kernels.
Flight control and estimation
Rigid-body mathematics, state estimation, simulation, and sensor fusion built methodically toward controlled autonomous systems.
Antsand and evidence
Structured experiments, telemetry, evidence dashboards, deployment tooling, and technical publishing that preserve the reasoning behind each result.
Current evidence, not future ambition
Every claim below links to a source artifact and carries an explicit engineering status.
Multi-family model bring-up
CKE supports multiple Qwen, Gemma, GLM, Nanbeige, and multimodal paths through explicit model-family contracts.
Generated C that fails closed
The v8 pipeline lowers model circuits to deterministic C and rejects missing kernels, routes, or numerical contracts instead of silently falling back.
Cross-runtime numerical parity
Layered scalar, ISA, and end-to-end gates compare CKE with PyTorch and llama.cpp and identify the first divergent operation.
CPU execution across real hardware
Current work spans AVX2, AVX-VNNI, AVX-512, AMX BF16, and ARM NEON, with measured support stated separately from work still being hardened.
Simulation and control foundations
Rigid-body mathematics, control systems, sensor models, and reproducible simulation are being developed before any autonomous-product claim.
A public engineering record
ShivasNotes publishes the derivations, architecture, debugging evidence, and limitations behind the work.
Autonomy is a chain of hardware and software contracts
These are research responsibilities and maturity targets, not a claim that a finished autonomous drone is currently available.
State and navigation
IN PROGRESS
PHYSICAL SYSTEM
- Rigid-body dynamics
- Actuator and vehicle models
- Physical constraints
SOFTWARE SYSTEM
- State estimation
- Sensor fusion
- Path and mission planning
Energy and endurance
IN PROGRESS
PHYSICAL SYSTEM
- Rotor and power models
- Battery and payload limits
- Measured test inputs
SOFTWARE SYSTEM
- Energy budget
- Mission constraints
- Range estimation
Safety and verification
IN PROGRESS
PHYSICAL SYSTEM
- Test constraints
- Benchtop and simulation rigs
- Failure instrumentation
SOFTWARE SYSTEM
- Deterministic simulation
- Parity and regression checks
- Fail-safe logic
Field sensing
PLANNED
PHYSICAL SYSTEM
- Payload interfaces
- Environmental sensors
- Edge compute
SOFTWARE SYSTEM
- Telemetry records
- Evidence pipelines
- Human review
Remote deployment
PLANNED
PHYSICAL SYSTEM
- Docking and charging
- Networking
- Serviceable hardware
SOFTWARE SYSTEM
- Long-horizon autonomy
- Fleet coordination
- Operational monitoring
The engineering record is part of the product
Each article turns a code change, mathematical question, or measured failure into a durable explanation.
Templates are circuit maps
How model-family structure becomes explicit compiler input rather than hidden runtime convention.
K-quants deep dive
How Q4_K, Q5_K, Q6_K, Q8_K, and mixed dot products move through real CPU kernels.
Prefill versus decode
Why one model becomes two CPU workloads with different matrix shapes, memory traffic, and execution plans.
Bounded engineering engagements
We take on well-scoped problems where a model, runtime, control system, or Linux CPU path is unsupported, slow, numerically wrong, or difficult to audit.
CPU AI investigation
Feasibility, performance, and first-divergence analysis for CPU inference.
Numerical parity
Runtime-versus-reference divergence localization, operation by operation.
C and Linux optimization
Kernel, memory-path, portability, and profiling work on real hardware.
Controls and simulation
Robotics mathematics, deterministic simulation, and control-system prototyping.
Antshiv Robotics builds intelligent systems we can understand all the way down, from mathematical kernels and CPU runtimes to simulation, flight control, and autonomous machines.