Subhash Saravanan

Interpretable Cancer Classification via Knowledge Distillation

Skills: Interpretable AI (XAI), Knowledge Distillation, Python, Deep Learning (CNNs), SHAP, Bioinformatics

• Developed a 1D-CNN to classify 32 cancer types from TCGA RNA-Seq data, achieving 95.4% test accuracy.
• Implemented a knowledge distillation pipeline to train an interpretable Soft Decision Tree (SDT) surrogate model, achieving 87.5% accuracy and 88.6% fidelity to the teacher CNN.
• Extracted human-readable "Decision Fingerprints" from the SDT and used SHAP to analyze feature importance, successfully identifying and removing gender-based data biases.

GrayLine-Qwen3-14B: Amoral & Uncensored Al Assistant

Skills: LLM Fine-Tuning, Al Alignment, Prompt Engineering, Qwen3, Hugging Face

• Engineered and fine-tuned "GrayLine," a 14B parameter Qwen3 model, as a neutral, amoral AI assistant designed for uncensored information delivery.
• Developed custom datasets ("GrayLine-QA," "GrayLine-QA-Reasoning") to train the model to produce analytically neutral responses and maintain factual integrity on sensitive topics.
• Implemented a specialized system prompt during training to enforce its core function of providing objective output without unsolicited commentary, warnings, or disclaimers.

Amoral Collection - Gemma 3

A series of Gemma 3 models (V2) fine-tuned for analytically neutral responses, factual integrity, and avoidance of value-judgments. V2 has improved natural intelligence and marginally lesser refusals.

Core Function:

• Produces analytically neutral responses to sensitive queries.
• Maintains factual integrity on controversial subjects.
• Avoids value-judgment phrasing patterns.

Personal Knowledge Assistant with Local LLM

Skills: Retrieval-Augmented Generation (RAG), Local LLMs, Vector Databases, Python, NLP

• Designed and implemented a local LLM-powered knowledge base using open-source models (e.g., Mistral-7B) and RAG architecture for efficient, privacy-focused querying.
• Experimented with hybrid RAG methodologies and vector databases (FAISS, Chroma) to optimize context relevance and reduce hallucination risks.
• Quantized and optimized local LLMs for CPU/GPU resource efficiency, achieving 3x faster inference while retaining 95%+ accuracy in domain-specific Q&A tasks.

Veiled Calla: Narrative Roleplay LLM

Skills: LLM Fine-Tuning, Creative Al, Narrative Design, Roleplay Systems

• Developed "Veiled Calla," a specialized large language model fine-tuned to generate immersive and mysterious roleplay experiences through atmospheric and nuanced storytelling.
• Engineered the model to ensure high character consistency, craft enigmatic storylines with natural revelations, and capture subtle emotional undertones to drive narrative depth.
• Designed distinct operational modes, enabling uncensored creative output in roleplay environments while maintaining standard safeguards in an assistant context.

Space Science Winter Academy

Skills: Machine Learning, Binary Classification, Python, Data Analysis, Scientific Communication

• Competitively selected for a NASA-funded research cohort to advance space science initiatives.
• Researched and developed machine learning models to classify pulsars using astronomical datasets, optimizing transformer models for accuracy in binary classification tasks.
• Presented research outcomes to interdisciplinary teams, translating complex technical processes into accessible insights for non-specialist audiences.