Engineering a Scalable Laboratory Infrastructure for Assembly Language Scaffolding: Design and Deployment of a Locally Optimized GenAI Assistant for the CODE-2 Educational Architecture

Abstract

The transition from high-level programming to assembly language constitutes a welldocumented pedagogical bottleneck in computer engineering curricula, particularly in largecohort laboratory settings where individualized scaffolding cannot scale. This paper presents the design, implementation, and technical evaluation of a locally deployable generative AI assistant engineered specifically for the CODE-2 educational processor architecture. The system is intended as laboratory infrastructure, not as a replacement for human instruction; its primary contribution is enabling scalable, privacy-preserving syntax scaffolding without dependency on cloud services or internet connectivity. A synthetic task bank of 50,000 instruction pairs was procedurally generated to cover the full CODE-2 curriculum. Three fine-tuning strategies were compared on a consumer GPU: Prompt Tuning, Low-Rank Adaptation (LoRA), and Full Fine-Tuning of a T5-Small encoder-decoder model. Full Fine- Tuning achieved 94.10% Exact Match on the held-out evaluation set, demonstrating that rigid assembly syntax requires full parameter adaptation. Post-training INT8 quantization via ONNX Runtime reduced inference latency by 69% (from 1,689 ms to 526 ms) on standard laboratory hardware (Intel i5, 8 GB RAM), with a precision loss below 1%. The resulting system operates entirely offline, precluding data exfiltration by design. The system is integrated into laboratory workflows as a supervised scaffolding tool, requiring mandatory emulator-based verification of all AI-generated code. Pedagogical implications are discussed as plausible benefits; no controlled learning-gains study is reported. The work demonstrates a replicable pipeline for building domain-specific language model infrastructure tailored to CPU-only educational environments.