I'm an AI builder and dual-degree student based in State College. I work across agent infrastructure, developer tools, and the last mile between a capable model and a product people can actually use.
Building where AI capability meets real human use.
STATE COLLEGE · SHARJAH · CHENNAI01 / EXPERIENCE LOG
I like the messy middle: taking capable technology out of the demo, understanding the human system around it, and shipping the thing that sticks.
Perplexity AI
Deploying Perplexity Computer with local businesses: finding operational bottlenecks, connecting tools and data, automating recurring work, and training owners to run the workflows independently.
Program detailsPerplexity AI
Led demos, workshops, and campus outreach for 500+ Penn State students, translating Perplexity into practical coursework, coding, and research workflows with 20%+ attendee activation.
KreupAI Technologies LLC
Built on DaftarAI, an LLM copilot for accounting workflows—tracing request flow across the codebase and improving prompts, tool execution, integration reliability, and failure handling.
Independent E-commerce Venture
Built an independent operation past $30K in revenue while managing sourcing, supplier relationships, pricing, fulfillment, and 35%+ profit margins.
02 / SELECTED WORK
Four systems across agent authorization, editor-native AI, inspectable model behavior, and cross-vendor GPU execution.
Runtime authorization for AI agents
A protocol for scoped user mandates, bound tokens, service metadata, and signed execution receipts—designed against impersonation, replay, excessive permissions, spoofing, and payment escalation.
Better prompts and voice inside the editor
Published extensions for structured prompt rewriting and Whisper-powered speech-to-text across VS Code and Cursor. Verified marketplace activity: 1,258 Prompt Enhancer downloads, 1,093 Speech downloads on Open VSX, and 1,496 Speech installs on VS Marketplace.
See how sparse routing becomes GPU work
An open-source systems lab that captures real OLMoE routing, compiles expert assignments into grouped GEMM workloads, and publishes reproducible A100 measurements through an interactive evidence explorer.
Making sparse MoE execution actually move
An experimental sparse-MoE runtime that groups token work by expert, with custom Triton kernels validated on an NVIDIA A40 and WGSL compute shaders running through Vulkan on an AMD RX 6800.
03 / OPERATING EVIDENCE
Computer Science at Penn State and Data Science at IIT Madras.
Built through independent e-commerce, supplier strategy, fulfillment, and margin discipline.
Verified downloads and installs across the three extension listings linked above.
AI demos, workshops, and activations across the Penn State community.