AmeforgeAmeforge

AI Research at Ameforge

Our laboratory focuses on computational efficiency and adaptive intelligence. We develop architectures designed for non-stationary environments where data distributions shift over time.

01
76%
Exact Match

SAM-G

Published

Structured Action Model

A compact dual-mode language model for offline structured action generation across ten application domains. The foundation behind Foros and GearCut. SAM-G is the only sub-200M model producing valid structured output zero-shot. New: SAM-G-CobraTooling (30.3M) scores 78.8% exact plan-match on agentic tool-call orchestration — see it live in the FOROS robotics demo at foros.run.

Foundation ModelHuggingFaceDual-modeOffline
100% valid structured output · 50× efficiency · CobraTooling 78.8% plan-match
02
73.6%
Aggregate Match

SAM-MM

Published

Structured Action Model — Multimodal

SAM-MM extends SAM-G with vision, audio, and a latent physics world-model. It perceives image frames, sound, and physical dynamics, then reasons and acts. Includes SAM-MM-Audio-Reasoning for sound-triggered control.

MultimodalVisionAudioPhysics World-Model
100% physics reasoning · 4 input modalities · Offline
03
1.12
Sharpe Ratio

EARCP

Published

Ensemble Auto-Regulative Coherence Protocol

A novel ensemble learning framework that dynamically weights multiple expert models (CNN, BiLSTM, Transformer) using both recent performance and inter-model coherence. Published on arXiv.

arXiv:2603.14651Open SourcePyPICKL License
Published 2026 · arXiv:2603.14651
04
85%
Entity Retention

SCLM

Proprietary

Stateful Coherent Language Model

SCLM adds persistent latent memory to any transformer language model, enabling coherence across long conversations without growing the context window. Built on the EARCP architecture (Encapsulation, Alignment, Revision, Coherence, Propagation). Available under a dual-license model.

EARCP ArchitectureDual LicensePyPI: saclmPatent Pending
Persistent latent memory · No context window growth
05
87.5%
Sparsity

Sparse Architectures

Prototype

Native Sparsity During Training

Research on integrating sparsity directly during training rather than applying post-training pruning. Our prototype achieves 87.5% attention sparsity with less than 2% performance degradation on TinyStories benchmarks.

PyTorchNVIDIA A100v15
Native sparsity during training · Less than 2% performance degradation
06
0
Forgetting events

MemoryBank

R&D

Persistent Domain-Segmented Memory

A persistent external memory system segmented by domain, enabling neural networks to learn continuously without forgetting previously acquired knowledge. Validated over 10 consecutive learning cycles across 3 distinct domains.

LPOLMemory4096 slotsContinuous LearningMulti-domain
Continuous learning across sequential domains · Zero catastrophic forgetting
07
99.9%
Coverage

NexusBPE

Internal Tool

Adaptive Multi-Domain Tokenizer Pipeline

A controlled tokenization pipeline designed for multi-domain training with precise data distribution management. Supports academic, code, mathematical, and narrative corpora with configurable composition ratios.

32k vocabularySentencePieceBPEMulti-domain
99.9% character coverage · Multi-domain corpus support

All metrics reported above are measured on our laboratory prototypes. Results may vary depending on hardware, framework version, and use case. Additional technologies are maintained confidentially pending patent applications. Full technical documentation is available to qualified investors and partners under NDA.