I'm a B.Tech Computer Science (Data Science) student at BVRIT Hyderabad, currently working as Tech Lead at Stud Entertainments, with prior internship experience at AWS, IBM, and Google. Outside of coursework and internships, I run an independent research track focused on making large-model architectures -- attention, memory, inference -- work under real hardware constraints instead of just on paper.
Most of what is here falls into four buckets:
- Memory and attention architectures for foundation models -- designing systems that scale sub-quadratically without giving up retrieval quality
- Consumer-hardware inference engineering -- squeezing large-model workloads onto a single 4GB-VRAM GPU through quantization, caching, and prediction tricks
- AI-native systems -- building operating system substrates where the AI model is the kernel, not an application running on top of one
- Developer tooling -- small, sharp tools (compilers, package managers, UI libraries) that solve one problem cleanly
I evaluate my own work adversarially before anyone else gets to. If a result looks too good, I go find out why before I publish it.
|
HGDM -- Hierarchical Gated Delta Memory 100% attention-free, byte-level sequence architecture with Hierarchical Temporal Decimation, Factored Bilinear State Highways, and a custom Triton fused scan engine (Nitro). Trained at 1B parameters on 1.46B tokens. Achieves O(1) inference memory -- VRAM footprint stays constant regardless of context length, verified empirically across 20x prompt scaling.
|
NCM -- Native Cognitive Memory Tensor-based episodic memory with state-conditioned retrieval for AI systems. Uses a four-field distance function -- semantic, emotional, state-conditioned, and temporal -- as the native memory subsystem for agent-style architectures. Ported to
|
Active research (not yet public):
- ZS-ISAB -- a method for scaling TabPFN's attention from quadratic to linear complexity via seeded anchor selection, an online softmax accumulator, and a zero-shot attention mask -- enabling 500K+ row inference on 4GB VRAM. Currently in iteration toward TMLR submission.
- PAS-Offload -- a consumer-GPU LLM inference engine combining CPU-side low-rank activation prediction, transposed column-major weight caching, and dynamic 2-bit weight slicing, benchmarked on an RTX 3050.
|
NOUS -- Cognitive Operating System A bare-metal AI-native OS in Rust for
|
A small, rule-based, deterministic compiler that takes a plain app description and generates native Web, Android, and iOS UI code without an LLM in the loop.
A Tkinter GUI for managing Python virtual environments and pip packages -- for when you do not want to remember venv flags.
|
|
Streamlit app powered by Gemini 2.5 Flash Image that turns one uploaded photo into platform-ready assets for Amazon, Flipkart, Instagram, Spotify, and more -- with in-place AI re-editing via chat instructions.
|
Lightweight React library for cinematic scroll-driven image sequences -- eager/lazy frame loading, TypeScript support, SSR-safe, and built-in accessibility (respects
|
Other repositories
- Windows-MCP -- contributing to an MCP server for computer-use automation on Windows
- ailang-py -- Python project for AI-native language utilities
- VoxMimic -- voice cloning and mimicry system
- storyboarding -- AI-assisted storyboard generation
- synthetic-life-system -- emergent behavior and synthetic agent simulation
- Event-Ticket-Booking-System, Buzzer, SmartFarm, Chaff -- earlier web and coursework projects
Core
ML and Infra
Cloud and Tools
- Iterating on ZS-ISAB toward TMLR submission -- resolving TabICL baseline verification and regenerating figures against the current seeded-anchor architecture
- Running a 91-dataset TabZilla benchmark against tree-based baselines
- Prototyping row-ordering strategies for fusing MiniMax Sparse Attention into a TabPFN-style tabular foundation model
- Continuing development on NOUS -- next milestone is a persistent NVMe driver replacing the simulated cold storage tier
- Open to research collaborations on efficient attention, memory-augmented architectures, and consumer-hardware ML systems