release: v0.3.0 — Production edge inference milestone

12 AxonML models (6 LSTM autoencoders + 6 GRU failure predictors) running
live inference on Raspberry Pi edge controllers across 5 buildings. Includes
autograd graph fixes, named_parameters for LSTM/GRU/StateDict, version bump
to 0.3.0, CHANGELOG and README updates with deployment details.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: AutomataControls

CHANGELOG.md

@@ -7,6 +7,58 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.3.0] - 2026-02-27
+
+### Milestone: Production Edge Inference
+
+AxonML models are running live production inference on 6 edge controllers (Raspberry Pi),
+monitoring HVAC equipment across 5 buildings. 12 models (6 anomaly detectors + 6 failure
+predictors) deployed via cross-compiled ARM binaries, each running at ~2-3 MB RSS.
+
+### Added
+
+#### Autograd Fixes (`axonml-autograd`, `axonml-nn`)
+- Fixed critical autograd graph-severing bug where `Variable::new()` was used for
+  intermediate results, creating leaf variables that blocked gradient flow
+- Fixed LSTM/GRU weight transpose operations (6 instances in `rnn.rs`)
+- Fixed `stack_outputs` in RNN/LSTM/GRU to use `unsqueeze` + `Variable::cat`
+- Added `CrossEntropyBackward` gradient function for proper backpropagation
+- Made `Variable::from_operation` public for custom gradient-tracked operations
+
+#### Tensor Operations (`axonml-tensor`)
+- `Tensor::cat(tensors, dim)` with `CatBackward` gradient function
+- `Variable::cat(vars, dim)` for autograd-tracked concatenation
+- `Tensor::sum_dim(dim, keepdim)` with `SumDimBackward` gradient function
+- `Variable::sum_dim(dim)` for autograd-tracked dimension reduction
+
+#### CUDA Backend (`axonml-core`, `axonml-tensor`)
+- CUDA matrix multiplication dispatch via cuBLAS GEMM
+
+#### Serialization (`axonml-serialize`)
+- Model save/load for production deployment (`.axonml` format)
+- StateDict extraction for weight export
+
+#### Production Edge Inference
+- Pure-tensor inference daemons (no autograd overhead) for ARM deployment
+- Cross-compilation pipeline for `armv7-unknown-linux-musleabihf` (static musl)
+- HTTP API endpoints (`/health`, `/api/inference/latest`) for integration
+- Rolling window buffers for time-series LSTM/GRU inference
+- PM2 process management for production uptime
+
+### Production Deployments
+
+| Building | Unit | Anomaly Model | Failure Predictor | Controller |
+|----------|------|---------------|-------------------|------------|
+| FCOG | Mechroom | Erebus (128K params) | Kairos (288K params) | 100.123.60.69 |
+| Warren | AHU-1 | Aether (32K params) | Moros (73K params) | 100.124.76.93 |
+| Warren | AHU-2 | Phanes (71K params) | Hecate (162K params) | 100.95.58.104 |
+| Warren | AHU-4 | Nyctos (32K params) | Cassandra (73K params) | 100.121.143.51 |
+| Warren | AHU-7 | Poseidon (32K params) | Triton (73K params) | 100.125.245.8 |
+| Huntington | Mechroom | Plutus (127K params) | Moira (288K params) | 100.73.201.107 |
+
+### Changed
+- Bumped version from 0.2.8 to 0.3.0
+
 ## [0.1.0] - 2024-XX-XX
 
 ### Added
@@ -120,7 +172,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Version History
 
+- **0.3.0**: Production edge inference — 12 models deployed across 6 controllers
 - **0.1.0**: Initial release with complete ML framework
 
-[Unreleased]: https://github.com/automatanexus/axonml/compare/v0.1.0...HEAD
-[0.1.0]: https://github.com/automatanexus/axonml/releases/tag/v0.1.0
+[Unreleased]: https://github.com/AutomataNexus/AxonML/compare/v0.3.0...HEAD
+[0.3.0]: https://github.com/AutomataNexus/AxonML/compare/v0.1.0...v0.3.0
+[0.1.0]: https://github.com/AutomataNexus/AxonML/releases/tag/v0.1.0

Cargo.lock

@@ -5,7 +5,7 @@
 # A PyTorch-equivalent machine learning framework written in pure Rust.
 # Provides tensors, automatic differentiation, neural networks, and more.
 #
-# @version 0.1.0
+# @version 0.3.0
 # @author AutomataNexus Development Team
 # @license MIT OR Apache-2.0
 # =============================================================================
@@ -38,7 +38,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.2.8"
+version = "0.3.0"
 edition = "2021"
 rust-version = "1.75"
 authors = ["AutomataNexus Development Team"]
@@ -52,24 +52,24 @@ readme = "README.md"
 
 [workspace.dependencies]
 # Internal crates
-axonml-core = { path = "crates/axonml-core", version = "0.2.8" }
-axonml-tensor = { path = "crates/axonml-tensor", version = "0.2.8" }
-axonml-autograd = { path = "crates/axonml-autograd", version = "0.2.8" }
-axonml-nn = { path = "crates/axonml-nn", version = "0.2.8" }
-axonml-optim = { path = "crates/axonml-optim", version = "0.2.8" }
-axonml-data = { path = "crates/axonml-data", version = "0.2.8" }
-axonml-vision = { path = "crates/axonml-vision", version = "0.2.8" }
-axonml-audio = { path = "crates/axonml-audio", version = "0.2.8" }
-axonml-text = { path = "crates/axonml-text", version = "0.2.8" }
-axonml-distributed = { path = "crates/axonml-distributed", version = "0.2.8" }
-axonml-serialize = { path = "crates/axonml-serialize", version = "0.2.8" }
-axonml-onnx = { path = "crates/axonml-onnx", version = "0.2.8" }
-axonml-quant = { path = "crates/axonml-quant", version = "0.2.8" }
-axonml-fusion = { path = "crates/axonml-fusion", version = "0.2.8" }
-axonml-profile = { path = "crates/axonml-profile", version = "0.2.8" }
-axonml-llm = { path = "crates/axonml-llm", version = "0.2.8" }
-axonml-jit = { path = "crates/axonml-jit", version = "0.2.8" }
-axonml-tui = { path = "crates/axonml-tui", version = "0.2.8" }
+axonml-core = { path = "crates/axonml-core", version = "0.3.0" }
+axonml-tensor = { path = "crates/axonml-tensor", version = "0.3.0" }
+axonml-autograd = { path = "crates/axonml-autograd", version = "0.3.0" }
+axonml-nn = { path = "crates/axonml-nn", version = "0.3.0" }
+axonml-optim = { path = "crates/axonml-optim", version = "0.3.0" }
+axonml-data = { path = "crates/axonml-data", version = "0.3.0" }
+axonml-vision = { path = "crates/axonml-vision", version = "0.3.0" }
+axonml-audio = { path = "crates/axonml-audio", version = "0.3.0" }
+axonml-text = { path = "crates/axonml-text", version = "0.3.0" }
+axonml-distributed = { path = "crates/axonml-distributed", version = "0.3.0" }
+axonml-serialize = { path = "crates/axonml-serialize", version = "0.3.0" }
+axonml-onnx = { path = "crates/axonml-onnx", version = "0.3.0" }
+axonml-quant = { path = "crates/axonml-quant", version = "0.3.0" }
+axonml-fusion = { path = "crates/axonml-fusion", version = "0.3.0" }
+axonml-profile = { path = "crates/axonml-profile", version = "0.3.0" }
+axonml-llm = { path = "crates/axonml-llm", version = "0.3.0" }
+axonml-jit = { path = "crates/axonml-jit", version = "0.3.0" }
+axonml-tui = { path = "crates/axonml-tui", version = "0.3.0" }
 
 # External dependencies
 thiserror = "1.0"

Cargo.toml

@@ -23,7 +23,7 @@ AxonML provides comprehensive PyTorch-equivalent functionality with 1076+ passin
 
 ## Features
 
-### Core (v0.2.8)
+### Core (v0.3.0)
 
 - **Tensor Operations** (`axonml-tensor`)
   - N-dimensional tensors with arbitrary shapes
@@ -423,7 +423,7 @@ Add Axonml to your `Cargo.toml`:
 
 ```toml
 [dependencies]
-axonml = "0.1"
+axonml = "0.3"
 ```
 
 ### Basic Usage
@@ -538,6 +538,23 @@ let v = t.slice_dim0(0, 1).unwrap();
 let n = t.narrow(1, 0, 2).unwrap();
 ```
 
+## Production Edge Deployment
+
+AxonML powers real-time predictive maintenance on HVAC systems across commercial buildings. 12 models (6 LSTM autoencoders for anomaly detection + 6 GRU failure predictors) run live inference on Raspberry Pi edge controllers, processing sensor data at 1 Hz.
+
+| Building | Unit | Anomaly Detector | Failure Predictor | Params | RSS |
+|----------|------|-------------------|-------------------|--------|-----|
+| FCOG | Mechroom | Erebus (LSTM-AE) | Kairos (GRU-FDD) | 416K | 2.5 MB |
+| Warren | AHU-1 | Aether | Moros | 105K | 2.1 MB |
+| Warren | AHU-2 | Phanes | Hecate | 233K | 2.4 MB |
+| Warren | AHU-4 | Nyctos | Cassandra | 105K | 2.1 MB |
+| Warren | AHU-7 | Poseidon | Triton | 105K | 2.1 MB |
+| Huntington | Mechroom | Plutus | Moira | 415K | 3.2 MB |
+
+**Stack:** AxonML training (CPU) → `.axonml` model files → cross-compiled ARM inference daemons (`armv7-unknown-linux-musleabihf`) → PM2-managed services on Raspberry Pi → REST API (`/api/inference/latest`)
+
+Each daemon runs pure-tensor inference (no autograd overhead), polls local NexusEdge for sensor data, maintains rolling time-series buffers, and exposes anomaly scores + failure predictions via HTTP.
+
 ## Architecture
 
 ```

README.md

@@ -5,7 +5,7 @@
 # Provides a unified CLI for project scaffolding, training, evaluation,
 # model management, and deployment of Axonml ML models.
 #
-# @version 0.1.0
+# @version 0.3.0
 # @author AutomataNexus Development Team
 # @license MIT OR Apache-2.0
 # =============================================================================
@@ -42,7 +42,7 @@ axonml-optim = { workspace = true }
 axonml-data = { workspace = true }
 axonml-vision = { workspace = true }
 axonml-serialize = { workspace = true }
-axonml-onnx = { path = "../axonml-onnx", version = "0.2.8" }
+axonml-onnx = { path = "../axonml-onnx", version = "0.3.0" }
 axonml-tui.workspace = true
 
 # CLI framework