San Francisco Bay Area
(+) ๐ ๐จ๐ฅ๐ฅ๐จ๐ฐ for latest and diverse perspectives on ๐๐, ๐๐๐ฎ๐๐๐ญ๐ข๐จ๐ง and ๐ ๐ฎ๐ญ๐ฎ๐ซ๐ ๐จ๐ ๐๐จ๐ซ๐ค! For speaking engagements, media, and business inquiries: [email protected]. Dhyey is driven by complex problems at the intersection of mathematics, computer science, statistics, and economics. His liberal arts foundation equips him uniquely to build innovative solutions, leveraging strategic thinking and deep technical expertise. A strategic thinker at heart, Dhyey excels in dissecting intricate problems and devising actionable solutions. He believes in the power of collaboration, cherishing partnerships with professionals from diverse backgrounds to produce outstanding results. He has built AI infra, models, and agentic product capabilities at ๐๐ข๐ง๐ค๐๐๐๐ง after short but formative stints as CDN engineer at ๐๐๐, and quantitative trader at ๐๐๐ฅ๐ค๐ฒ๐ซ๐ข๐. At LinkedIn, he delivered major LLM inference wins with speculative decoding, and significant GPU savings while pioneering LinkedIn's first Rust-based model serving engine for PyTorch Embedding Based Retrieval (EBR) models. He also integrated MCP into enterprise workflows that contributed to scalable model onboarding across Azure OpenAI and Kubernetes. He even proposed strategic ideas directly to CEO Ryan Roslansky, leading to cross-functional collaborations. He graduated from ๐๐ฆ๐ก๐๐ซ๐ฌ๐ญ ๐๐จ๐ฅ๐ฅ๐๐ ๐ with a ๐ญ๐ซ๐ข๐ฉ๐ฅ๐ ๐ฆ๐๐ฃ๐จ๐ซ in Computer Science, Mathematics, and Statistics, earning 2 ๐๐ฎ๐ฆ๐ฆ๐ ๐๐ฎ๐ฆ ๐๐๐ฎ๐๐ theses, along with induction to ๐๐ก๐ข ๐๐๐ญ๐ ๐๐๐ฉ๐ฉ๐ and ๐๐ข๐ ๐ฆ๐ ๐๐ข honor societies. His mathematical work includes building a Lean4-based machine-assisted proof framework for chip-firing and Graphical RiemannโRoch alongside a Python package for simulations (chipfiring), while his statistics thesis led to โccrvam,โ an open-source Python package for discrete copula modeling. He also completed six graduate-level math and statistics courses at ๐๐จ๐ฅ๐ฎ๐ฆ๐๐ข๐ in one semester and continues advanced computational mathematics studies at ๐๐ญ๐๐ง๐๐จ๐ซ๐. Outside the professional sphere, adventure calls out to Dhyey. Whether it's traversing new terrains, perfecting his archery shot, or engaging in a spirited badminton match, he's always up for a challenge. Other adventurous interests of Dhyey include jetskiing, ATV quad biking, and bungee jumping. **Opinions, posts, and views expressed here are solely Dhyey Mavani's own and do not reflect that of his employers.**
2025-2026: - Building the next generation of digital work AI agents, directly with LinkedIn's Chief AI Officer, President of Engineering, and LinkedIn's CEO. - Founded & Led a pod of 3 directly with Chief AI Officer, published a research paper on stability margins for LLM robustness (coined "Support Tokens"). - Youngest MLE working daily alongside Distinguished Engineers; Dabbling across the board: Influencing product, GTM, and governance across context, tools, and agent infra. - Promoted to Highly Confidential capacity, and given access to tented access to broader MSFT + OpenAI resources. - Secured override to traditional policies to build-my-own role after emailing the C-suite! 2025 H1: - Optimized LLM inference in Hiring Agent (LangGraph); Increased throughput by 4x; Slashed latency by 66% with speculative decoding. - Built first-ever MCP integrations with Azure OpenAI deployment infra to improve model onboarding and k8s quota management. - Built SOTA fine-grained observability into the production vLLM and SGLang based hybrid serving stack with a centralized dashboard. - Recruited, Interviewed and Mentored Masters & PhDs on AI assisted coding and LLM serving specifics. Summer 2024: - Boldly pitched 3 novel product ideas to the CEO, Ryan Roslansky as an intern, and that led to further discussions with product leaders across 3 business units for formulating them for the future B2B2C strategies at LinkedIn. - Built & integrated an end-to-end Rust-based low-latency inference engine for Torch. Developed first production-ready PyTorch EBR kNN model with custom-filtering kernels. - My work was being actively used to serve 5+ business use-cases leading to P99 inference latency reduction from 4 ms to 2 ms, and P99 tail data ingestion latency reduction from 70 ms to 7 ms. This also increased throughput by 4x, leading to GPU savings of ~ $XX million.
ccrvam - Python package on a novel discrete copula method (30k+ downloads in a few months) - summa cum laude honors chipfiring - Python pakage on an open graph / number theory problem (7k+ downloads and industry adoption) chip-firing-with-lean - First-ever formalization of chip firing dynamics in Lean4 (addressed a gap in Microsoft developed language); summa cum laude honors QuantileFlow - state of the art performance quantile calculations in service logs (beating Datadogโs Python; 6k+ downloads) RBlocks - block-based programming interface for R; deployed in classrooms for accessibility; addressed gap in Googleโs Blockly
Developed a Monte-Carlo simulation-based strategy for theo-value estimation and shadowed quants/traders in treasuries and equity options desks. Gained hands-on experience in the execution and risk management of delta and gamma hedging market-making strategies through intensive mock sessions.
Populated a metrics dashboard, pushed latency optimizations to production at AWS CloudFront; Proposed & documented use-case of AutoML to analyze change-propagation/caching service logs to predict future outages
โก Columbia x Citadel Trading Challenge based on Auction Theory (New York, Apr 2023) โก Akuna Options 201 course (Virtual, Apr 2023) โก Schonfeld Early Engagement Summit (New York, Apr 2023) โก Amplify x Citadel Trading Challenge (New York, Feb 2023) โก D.E. Shaw & Co. Connect Event (New York, Sep 2022) โก D.E. Shaw & Co. Latitude Fellowship (New York, Aug 2022) โก Citadel Securities Invitational Terminal (Virtual, Aug 2022) โก Citadel Securities Invitational Datathon (Virtual, Jul 2022) โก Hudson River Trading BIPOC Tech Summit (Virtual, Jun 2022) โก Goldman Sachs Insights Program (Virtual, Apr 2022 - Jun 2022) โก Jane Street SEE Quant Research & Trading (New York, May 2022) โก Jane Street FOCUS Software Engineering (New York, Mar 2022) โก SIG Freshman Discovery Program (Virtual, Apr 2022)