ANNA NEXUS LABS LLC / OPE3.27N NEURAL FORGE

ANNA Neural Intelligence

ANNA Neural Intelligence is a hardware-aware neural-network engineering project developed by Anna Nexus Labs LLC. It explores neural-network training, inference performance, optimization, quantization, and the relationship between machine-learning architectures and computing hardware.

Locally trained and CPU measured. Evidence recorded October 13, 2026. Owner approved public release on October 13, 2026. Complete source and measured evidence are publicly available on GitHub. The source was developed with AI assistance and tested locally.

A trained network alongside ANNA

ForgeCNN is an original, standalone PyTorch convolutional model trained from random weights on UCI Optical Digits. ANNA 1.3 remains the separate multi-agent orchestration system. Its live provider experiment is unapproved, with $0 authorized spending. Connecting this classifier to ANNA is future work.

Model architecture

38,282 trainable parameters. Two convolutions with ReLU, one pooling layer and a 64-unit classification head produce ten logits. Estimated convolution and linear work: 116,352 multiply-accumulates per image, excluding other operations.

ForgeCNN layers and tensor shapes from normalized 8 by 8 grayscale input through two convolutions, pooling, flatten and ten output logits — The dense hidden layer owns most parameters. The second convolution owns most arithmetic work.

Training and held-out evaluation

3,249 training images, 574 validation images and an official test file with 1,797 images. Normalization fits only training data. AdamW trains for 40 fixed epochs with batch size 64 and learning rate 0.002. The reference checkpoint minimizes validation loss at epoch 14. Test data does not select weights or training settings.

Three fixed seeds achieved 97.61%, 97.33% and 97.38% test accuracy. The predefined reference seed is 369. Repeating it reproduced identical selected weights and test metrics in this CPU environment.

Measured training and validation loss and accuracy over 40 epochs — Checkpoint selection uses validation loss, so the selected epoch can differ from the final epoch.

FP32 reference confusion matrix for the held-out UCI test set, with digit classes zero through nine

Precision and storage tradeoffs

Official held-out test accuracy and serialized model state
Variant	Accuracy	State size
FP32	97.61%	156,412 bytes
FP32 channels-last	97.61%	Same logical weights
INT8 x86	96.72%	47,854 bytes

INT8 calibration uses 512 training images. The quantized state is 3.27 times smaller, but accuracy drops by 0.89 percentage points. Serialized state excludes the optimizer and runtime. INT8 checkpoint reload reproduced identical logits locally.

Measured CPU inference

Xeon Platinum 8573C CPU in a shared Linux container, eight-core quota and an 8 GiB memory limit. PyTorch 2.5.1 CPU with Python 3.12. No GPU or custom Tesla hardware was available.

36 measured conditions compare FP32, channels-last FP32 and INT8 at batches 1, 8, 32 and 128 with 1, 2 and 4 intra-op threads. Each condition uses 30 warmups and five repetitions of 100 timed forwards. Preallocated synthetic input is used for timing; real held-out data is used for accuracy.

One-thread subset. Full matrix and variability available below.
Variant	Batch	Median ms	p95 ms	Images/s
fp32	1	0.0660	0.3932	7,764
int8	1	0.2695	0.9936	2,606
fp32_channels_last	1	0.0553	0.2499	11,301
fp32	8	0.2724	0.9089	21,430
int8	8	0.2818	0.8078	20,902
fp32_channels_last	8	0.1769	0.5172	33,142
fp32	32	1.4586	3.2477	17,786
int8	32	0.3714	0.8605	72,989
fp32_channels_last	32	0.3507	1.0294	69,056
fp32	128	2.0358	2.7771	60,340
int8	128	0.9667	1.6921	121,799
fp32_channels_last	128	1.0718	1.9190	107,984

Measured one-thread forward latency and mean-based throughput versus batch size for FP32, channels-last FP32 and INT8

INT8 is slower at batch one and faster than contiguous FP32 at larger one-thread batches in these observations. Channels-last also improves several conditions. Shared-host timing variability limits performance conclusions. These ratios do not transfer automatically to another platform.

Verification and limitations

Locally verified: 29 neural tests, full training, checkpoint restoration, held-out metrics, quantization, CPU benchmarks and profiles.
Simulated: only ANNA's separate retained orchestration simulations; neural training results here are real local measurements.
Planned: larger datasets, additional hardware and separately reviewed ANNA integration.
Unmeasured: GPU and Tesla accelerator performance, energy, sustainable memory bandwidth and production behavior.
Public release: owner authorized. These are locally measured CPU results; external reproduction remains open. No employer endorsement or certification is claimed.

Process peak RSS includes Python and PyTorch; it is not isolated activation memory. First-forward timing differs from full process startup. Strict FP32 training stays enabled, while INT8 inference has a scoped deterministic-enforcement exception for a pinned PyTorch kernel limitation.

Reproduce and inspect

Engineering report · Layer specifications · Benchmark methodology · Dataset and license · Complete benchmark CSV · Measurement summary JSON

GitHub repository: Christopher4U1978/anna-neural-intelligence. Source, raw samples, model checkpoints and demonstration materials are in the review package. No placeholder repository URL is presented.

python -m unittest discover -s tests -v
python -m anna_neural.cli train --out checkpoints/my-run

The UCI dataset has a CC BY 4.0 license. Alpaydin and Kaynak, Optical Recognition of Handwritten Digits (1998). Small digit classification does not establish production vision performance.