I'm Dushmilan — a first-year Computer Science undergraduate at the University of Peradeniya, Sri Lanka, with a passion for AI infrastructure engineering and audio AI research.

• CodeCoach-AI — An open-source educational platform for ~400 CS students at the University of Peradeniya, featuring an integrated AI chat panel. Transitioning from user-provided API keys to a self-hosted fine-tuned Gemma 2B model for sustainable, scalable deployment.
• Research at MARC — Contributing to the MultiDisciplinary AI Research Centre at the University of Peradeniya, exploring cutting-edge applications of machine learning.

• Java — Full-stack development with a focus on backend systems and networking
• Fine-tuning & LLM Deployment — Hugging Face LLM course, model quantization, and scalable inference
• Advanced Audio AI — mel-filterbanks, speaker embeddings (ECAPA-TDNN), speech tokenization (wav2vec2, VQ-VAE), and generation using frameworks like SpeechBrain and PyAnnote
• Differential Equations & Dynamical Systems — Deepening mathematical foundations for advanced AI research
• Test-time Compute & Low-resource NLP — Addressing efficiency and accessibility in language models

My long-term goal: Direct PhD admission into top-tier Audio Reasoning and Large Audio-Language Model (LALM) research labs.

Core focus: How can LLMs better understand and generate speech by capturing speaker nuances, emotions, and natural pronunciation using exclusively audio signals?

Key domains:

• Audio embeddings & speaker recognition (VoxCeleb, wav2vec2, CLAP)
• Speech tokenization & generation
• Low-resource & multilingual audio processing
• PINN training for acoustic modeling
• Ethical & regulatory frameworks for audio AI

Languages: Python, JavaScript, Java (learning)
ML/Audio: TensorFlow, PyTorch, SpeechBrain, PyAnnote, Hugging Face
Web: React, Node.js, FastAPI
Tools: Git, Docker, Figma, Notion
Databases: PostgreSQL, MongoDB

• Audio AI & speech processing projects
• Full-stack ML platforms for education or research
• Open-source contributions in LLM inference & fine-tuning
• Research on test-time compute and efficient model deployment

• Building full-stack educational AI platforms from idea to production
• Self-directed CS & statistics self-study beyond formal coursework
• Sri Lankan tech ecosystem — internships, startup culture, opportunities
• Audio ML fundamentals — embeddings, speaker systems, tokenization
• Bridging academia and industry in AI research

• LinkedIn: Dushmilan Jeyanathan
• GitHub: github.com/Dushmilan-Jeyanathan
• Email: Dushmilan05@gmail.com

• Built a production e-commerce PWA (Twinkle-Hearts) integrated with WhatsApp Business API before turning 19
• Achieved 87.5% computational reduction in voxel-to-point-cloud processing for medical imaging
• Publicly documenting my learning journey — from Java systems programming to audio AI research
• Self-studied entire CS courses beyond university curriculum; currently optimizing a 4-year PhD prep roadmap
• Physics + Math + CS triple major focus; leveraging physics intuition for ML research

⭐ If you're interested in audio AI, LLM infrastructure, or building AI tools for education, let's connect!