Hi there, I’m Aditya.
I study how LLMs and transformers work, not the ones from the movies, unfortunately. The other kind!
What’s actually wired together under there, how information moves through it, how it can be nudged into behaving differently, and the math underneath all of it that the marketing conveniently skips.
My interest sits squarely in the field of mechanistic interpretability, simply said: less “does the model perform well,” more “what is it actually doing in there, and can I catch it in the act.”
Away from a terminal, it’s usually music on loop, an ongoing and only mildly one-sided appreciation for art, and philosophical rabbit holes I have no business going down.
This site is the paper trail. The parts that worked, the parts that spectacularly didn’t, and the slow, unglamorous middle where most of the actual learning happens. Projects, write-ups, papers, half-formed ideas that eventually became something. Read in order or don’t; either way, it’s all still here.
Machines converge. But humans? They transcend.
Aditya Pratap Singh
Current focus
Mechanistic Interpretability
• Attention & Circuit Analysis
• Activation Steering and Eval awareness
• Applied ML Systems Rigorous, Reproducible Experimentation
Below are selected systems and research-oriented projects that reflect my current technical focus.
Selected Projects
Tesseract v1 — Text-to-3D Mesh Generation Engine
August,2025
- Built a diffusion-based system for text-to-3D mesh generation with a reproducible, production-oriented inference pipeline.
- Focused on system design: stateless execution, device-aware fallback, modular components, and config-driven experimentation.
Reddit-Persona — LLM-based User Persona Generation
July,2025
- Developed a production-grade LLM system to analyze Reddit user activity and generate structured, UX-oriented personas.
- Implemented chunked inference, modular configuration, and dual interfaces (CLI and Streamlit) to balance cost, scalability, and usability.
MÍMIR — Research-Oriented LLM System (Early Stage)
2025 – Present
- Designing a research-oriented LLM system to study retrieval-augmented reasoning and parameter-efficient adaptation under realistic system constraints.
- Emphasizes reproducible evaluation and system-level trade-offs relevant to long-running ML services, rather than application-level demos.
Ongoing Work
●LLM Compression & Interpretability(paper in preparation)
Manuscript in preparation
Studying how parameter-efficient fine-tuning and compression techniques (LoRA, quantization, pruning) alter internal representations and attention dynamics in transformer models.
●MÍMIR — Cognitive DevOPS Diagnostic LLM System(Early Stage)
Active development
A research-oriented, production-grade LLM system for DevOps and SRE incident reasoning, emphasizing retrieval grounding, structured reasoning, and reproducible evaluation.
Leadership & Community
President,
Advait
September,2025 - Present
Member Count : 300 +
I lead Advait, a 300+ member student-led AI community focused on research-oriented machine learning, systems engineering, and applied AI development.
My role spans both technical leadership and organizational execution, including:
- Designing and driving research, engineering, and project-based initiatives across LLMs, computer vision, and ML systems.
- Organizing technical talks, workshops, and internal study groups, ranging from neural networks fundamentals to production-grade ML practices.
- Mentoring teams on end-to-end system building, emphasizing reproducibility, modular design, and real-world constraints.
- Coordinating cross-functional operations: event management, speaker outreach, sponsorship communication, public relations, and community growth.
- Overseeing technical direction, team structure, and execution quality, while managing social media presence and external communications.
Advait serves as a platform for translating academic curiosity into disciplined engineering practice, and for cultivating a culture centered on rigor, collaboration, and long-term skill development rather than short-term hype.