Mithravardhan P N

Mr. Mithravardhan P N


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
1 September 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
APAkshay Paranjape
Akshay Paranjape
MNMithravardhan P N
Mithravardhan P N
PSPriyanshu Singh
Priyanshu Singh
YMYelisetti Ssr Murthy
Yelisetti Ssr Murthy
Deploying large language models (LLMs) on-premise for domain-specific industrial applications, such as coordinate measuring machines (CMMs), presents distinct challenges and opportunities. This work presents a practical pipeline for hosting LLMs on-premise, using model distillation to achieve fast inference, data privacy, and in-house control. CMM-specific student models meet the need for offline, near-real-time diagnostic support and error-log interpretation while keeping proprietary measurement routines and quality data on the factory floor, away from cloud-based application programming interface (API) exposure. We describe a five-stage methodology covering teacher/student model selection, domain-specific data curation, knowledge distillation with feature transfer, post-distillation supervised fine-tuning, and model compression via quantization. The feature-transfer step aligns the student’s final-layer hidden states with the teacher’s through a single learnable linear projection P ∈ RdT ×dS , which maps the student hidden dimension dS to the teacher dimension dT and is optimized jointly with the student under a mean-squared-error objective. Training data were curated through a hybrid strategy that combines the parsing of International Organization for Standardization (ISO)-compliant calibration records with 3,900 prompt–response pairs assembled via multi-model teacher guidance. The distillation objective is a multi-objective loss that combines cross-entropy, KL divergence over a shared top-k logit support, and hidden-state alignment. Experimental results show that the distilled 0.5B-parameter student models retain 93.4–96.1% of teacher (7.5B–8B parameter) performance in full precision under our composite evaluation metric, while reaching up to a 32× memory-compression ratio after quantization with an accompanying inference speedup of roughly 1.6×. The final quantized models run at 34–38 ms/token (time to first token) on consumer-grade hardware, indicating that compact specialist models are economically practical for privacy-sensitive manufacturing environments. We position the work as a systems and domain-adaptation contribution rather than a new distillation algorithm, and we discuss its limitations candidly.