TsugiAI
Algorithms and adaptation. Continual learning, signal-quantized optimization, model behavior under hardware constraints.
-
Continual Learning Without Catastrophic Forgetting
Patent pending -
Signal-Quantized Optimization for Edge Inference
Patent pending -
Adaptation Methods for Foundation Models on Constrained Hardware
Patent pending
TsugiAI
The algorithms lab. Continual learning, signal-quantized optimization, and the parts of model behavior that survive contact with constrained hardware. Output is methods, ablations, and write-ups.
- Patent pending
Continual Learning Without Catastrophic Forgetting
Methods for sequential task learning where prior capability is preserved without rehearsal. Active research line with workshop submissions in preparation.
- Patent pending
Signal-Quantized Optimization for Edge Inference
Sign-quantized gradient methods for adaptation, alongside renormalization-derived multi-resolution and coupled-map-lattice gradient smoothing optimizers for non-convex objectives at constrained per-step compute.
- Patent pending
Adaptation Methods for Foundation Models on Constrained Hardware
Parameter-efficient fine-tuning techniques targeted at hardware below datacenter scale: phones, edge boxes, single-GPU workstations.
- Research
Routing Strategies for Mixture-of-Experts on Edge Silicon
Expert-selection policies that respect on-die memory layout and activation-tensor traffic patterns. The router is itself an optimization problem.
- Research
Calibration of Generative Models for Cinema-Grade Output
Color, motion, and frame-rate fidelity targets for video-generative models intended to share a delivery pipeline with traditionally captured cinema.
- Open source
Open-source SDKs: pip install tsugi
Apache-2.0 SDKs covering cross-rack training resilience and K-of-N LoRA continual learning.
K-Pool LoRA selects an active adapter slot via a frozen-encoder Gaussian-mixture router that is never updated by the language-modeling gradient. Composition with a sign-quantized active-slot optimizer and a fragility-aware eviction policy keeps a fixed-size K-pool stable across five sequential domains at both 1.5B and 7B scales. Replay buffers are not required, which matters under HIPAA, FINRA, and ITAR regimes.
Two orthogonal lines. The Signum-class sign-quantized optimizer halves optimizer-state memory against Adam-class baselines without measurable mean-quality regression at the empirical learning-rate plateau. The renormalization-derived family covers multi-resolution refinement (MRRO) and coupled-map-lattice gradient smoothing (CMLGS) at O(D) per step, where the CMA-ES family becomes infeasible above roughly D=5000 dimensions.
The constraint set is real: 16 to 32 GB of DRAM on consumer single-GPU systems, 8 to 16 GB on edge boxes, sub-watt power budgets at the bottom. K-Pool LoRA covers the continual-learning corner of this surface. The broader research line characterizes the parameter-efficient methods that survive contact with shipping memory hierarchies and on-die activation budgets.
Datacenter MoE routing assumes uniform interconnect bandwidth and ample HBM. Edge silicon has tiered memory (on-die SRAM, LPDDR, NAND) with order-of-magnitude bandwidth disparities. The router becomes its own optimization problem: minimize cumulative activation transfer subject to a quality target, with awareness of which experts are resident at which level.
Diffusion-class video generators trained on web-scraped datasets exhibit color-space drift, motion-vector inconsistency, and frame-rate quantization artifacts that pass through consumer encode chains but fail under cinema-grade evaluation (PQ HDR color volume, sub-pixel motion accuracy, 24p and 48p frame-rate fidelity). The line characterizes those drift modes and proposes calibration objectives aligned to cinema-grade encode targets.
The lab’s shipping surface. The unified tsugi meta-package wraps two Apache-2.0 SDKs: a cross-rack distributed-training resilience toolkit and a K-of-N LoRA continual-learning toolkit, both installable from PyPI and developed in the open.
