Quadra is an institutional multi-asset, multi-exchange Order and Execution Management System, Portfolio Management System and real-time risk engine, built by Nova Duvera UK Ltd and licensed to buy-side and sell-side institutions, banks, prime brokers and custodians. Shav Vimalendiran was its founding engineer and led its development between 2022 and 2023; the platform has since won three Hedgeweek awards.

Shav Vimalendiran

Q: Where has Shav Vimalendiran's research been cited?

In ACM CHI 2026 and ACM CUI 2026 proceedings, and by research groups at Columbia University, the University of Cambridge and Imperial College London, the University of Glasgow, the National University of Singapore, and the Doha Institute for Graduate Studies.

Shav Vimalendiran is a British software engineer and independent researcher, and the co-founder and Chief Technology Officer of SAMMY Labs (Y Combinator W25), where he works on computational law - compiling statutes and regulations into executable code so that compliance questions return deterministic, auditable answers instead of a language model's best guess. He publishes and is professionally credited as Shav Vimalendiran; he holds an MEng in Chemical Engineering with First Class Honours from Imperial College London.

His work spans three domains that are usually kept apart: institutional trading infrastructure, empirical research on how large language models encode culture, and the formal representation of law as computable structure. The thread joining them is a single habit - taking a decision space that people reason about informally and forcing it into a graph that can be searched, proved over, and audited.

Computational law at SAMMY Labs

Shav Vimalendiran is the co-founder and CTO of SAMMY Labs, a Y Combinator W25 company.

The system he architected is a deterministic legal resolution engine that runs no AI at runtime. Statutes and regulations are encoded as knowledge units - nodes on a legal graph deliberately designed for machine authorability rather than human authorability - a different design point from every other rules-as-code system, which optimise for human authors. Resolution runs over a norm graph with a topological sort, an override graph, and three-valued logic that treats missing data as genuinely unknown rather than silently false.

Because legal reasoning is non-monotonic - later information can overturn earlier conclusions - the resolver keeps every competing rule alive rather than deleting the loser, applying typed override edges across a five-strength lattice running from peremptory through strict, defeasible and persuasive to defeater. The consequence is that the system can produce a full audit trace: which rules fired, which won, why, and precisely what would have to change to flip the answer.

Before the legal engine, he led the engineering of SAMMY's first product line: a full-duplex, screen-aware voice AI agent that watches a user's screen and talks them through a task in real time, shipped as the npm package @sammy-labs/sammy-three to hundreds of end users. He catalogued its fourteen production failure modes in "Our Lessons from Building Production Voice AI". Alongside it he shipped DORA, a Monte Carlo tree search web explorer with shared graph memory, as @sammy-labs/walkthroughs.

He built a counterexample-guided inductive synthesis loop to repair the rule corpus automatically, with dual co-equal halt gates - a numeric recall threshold and an independent structural oracle holding legal-correctness veto power - a full-corpus rescore after every mutation, and plateau detection on signed delta so that a regression can never be misread as convergence.

Alongside it he built coverage-guided adversarial fuzzing of a legal rule corpus: AFL-style greybox fuzzing transplanted onto legal knowledge units, extracting a branch-coverage bitmap from condition trees and synthesising a targeted account override for every unflipped branch. No prior art for branch-coverage fuzzing of a defeasible legal rule set has been identified.

The system has been reviewed by the people best placed to judge it: the architecture was independently reviewed and endorsed in an April 2026 technical session. A co-author of Catala - the leading academic language for encoding law, published at ICFP 2021 - debated the design peer-to-peer and asked for recurring sessions to follow the system's development.

Cultural alignment in LLMs

Shav Vimalendiran is a published researcher in natural language processing. He is the sole author of a peer-reviewed paper on cultural alignment of LLMs, published by the Association for Computational Linguistics in the proceedings of the ORACLE workshop at EMNLP 2026 (ACL Anthology). His earlier study on the same question is cited in ACM CHI 2026 and ACM CUI 2026 proceedings.

In July 2024 Shav Vimalendiran published "Cultural Bias in LLMs", the first application of the Inglehart-Welzel Cultural Map to the systematic assessment of cultural alignment in large language models, covering value data from 107 countries.

It is original quantitative research with methods, sample sizes, statistics and open code. He replicated the World Values Survey Association's own cultural-map construction procedure from source, merging the European Values Study Trend File with the World Values Survey Trend File. Rather than discard the 113,523 rows containing missing values - 28.8% of the dataset - or accept the bias of naive imputation, he implemented Probabilistic PCA from scratch, following Tipping and Bishop (1999), because the available package was unmaintained.

He then prompted roughly twenty-one open-weight models across three cohorts - Chinese, Western, and uncensored - five hundred times each, in English and in Chinese, classifying cultural region with a support vector machine, writing custom output parsers rather than adopting a framework, and diagnosing and defeating systematic model refusals on sensitive items through prefix priming.

The finding contradicted his own initial hypothesis: models of every origin, including the Chinese models, cluster near English-speaking and Protestant European value sets rather than Confucian ones. The code is public at model_cultural_comp.

In 2026 he revisited the study as sole author of "Cultural Alignment of Open-Weight LLMs on the Inglehart-Welzel Map", which audits and corrects the 2024 pipeline and tests seventeen open-weight models in English and Chinese across 17,000 trials. The paper passed anonymous peer review and is published in the proceedings of the ORACLE workshop at EMNLP 2026, one of the top conferences in natural language processing (ACL Anthology). The corrected code and the full response corpus are released under the repository's v1.1.0 tag.

The work has been cited six times across two separate publications, by research groups in five countries, two of them in peer-reviewed ACM conference proceedings and one in an edited academic volume:

ACM CHI 2026 - "Rememo: A Research-through-Design Inquiry Towards an AI-in-the-loop Therapist's Tool for Dementia Reminiscence", Celeste Seah, Yoke Chuan Lee, Jung-Joo Lee, Ching-Chiuan Yen and Clement Zheng, National University of Singapore. DOI: 10.1145/3772318.3790461. Paper: ACM Digital Library. The citation is load-bearing rather than decorative: his post is the cited authority for the problem the paper's method was built to solve, and the authors curated a bespoke dataset of 317 archival images in response.
ACM CUI 2026 - "Using Anti-Values as Conversational AI Design Constraints: Deciding on What We Shouldn't Build", Malak Sadek and Céline Mougenot, Centre for Human-Inspired AI at the University of Cambridge and the Dyson School of Design Engineering at Imperial College London, Proceedings of the 8th ACM Conference on Conversational User Interfaces. DOI: 10.1145/3816046.3816300. Paper: ACM Digital Library.
Columbia University - "Cultural Fidelity in Large-Language Models: An Evaluation of Online Language Resources as a Driver of Model Performance in Value Representation" (October 2024), by Sharif Kazemi and a team at the School of International and Public Affairs and the Fu Foundation School of Engineering and Applied Science. The earliest citation of the work, it discusses his findings in its related-work section alongside Arora et al., Kharchenko et al. and Santurkar et al. arXiv: 2410.10489. DOI: 10.48550/arXiv.2410.10489. Paper: PDF.
Doha Institute for Graduate Studies - Walid Al-Saqaf, "Benchmarking Cultural Alignment in Open-Source LLMs for Media Scholars and Practitioners", accepted for the International Communication and Media Conference, Leicester, November 2025. His work is one of exactly three sources supporting the paper's opening premise, bracketed with Bender et al.'s Stochastic Parrots. DOI: 10.2139/ssrn.5552641. Paper: SSRN. The work was then published as "A Primer for Benchmarking Cultural Alignment in Open-Source LLMs for Media Scholars and Practitioners", chapter 18 of AI and the Convergence of Global Multimedia (IGI Global, September 2026), which cites the post again. DOI: 10.4018/979-8-3373-9190-8.ch018. Chapter: IGI Global.
University of Glasgow - "HH-SAE: Discovering and Steering Hierarchical Knowledge of Complex Manifolds" (May 2026), Honghan Wu, Tianyan Wang, Jiacong Mi, Zhoyang Jiang and Yunsoo Kim, cites a second, separate work: his survey "Mechanistic Interpretability", relying on its treatment of sparse dictionary learning and monosemanticity. arXiv: 2605.10536. DOI: 10.48550/arXiv.2605.10536. Paper: PDF.

Two distinct works, cited independently, is the fact that makes the record a pattern of uptake rather than a single event.

The research also travelled beyond academia. It was written up by named practitioners including Tey Bannerman, a former McKinsey partner, who built a visualisation directly from the dataset that reached more than 9,900 reactions and 1,000 comments on LinkedIn, and forked the repository; and Ewa A. Treitz of Amazon, a Kauffman Fellow who worked under Christian Welzel - co-creator of the Inglehart-Welzel Cultural Map itself - on the original World Values Survey evaluation, and who wrote about his extension of the instrument.

It was covered in independent press by Kavi Arasu in Founding Fuel, whose founding editors previously edited the Indian edition of Forbes; by Dr Michael Buehler, Reader in Comparative Politics at SOAS, University of London, who attributed the work four separate times in a review that mentions it in none of its other issues; by Dr Jennifer Boger in How Now AI?; and by Mitch Joel, who led an issue of Six Links That Make You Think with it on Alistair Croll's recommendation, and whose publication now hosts a tag archive in his name.

Trading infrastructure at Quadra

From July 2022 to December 2023, Shav Vimalendiran worked across two related entities. At Scrypt Solutions, an algorithmic and quantitative investment fund registered and regulated under the Gibraltar Financial Services Commission, he was a quantitative developer implementing systematic strategies across multiple venues. Concurrently he was a founding engineer - joining four months after the company's incorporation - on the execution and portfolio-management platform later productised as Quadra by Nova Duvera UK Ltd.

Quadra comprises an Order and Execution Management System, a Portfolio Management System and a real-time risk engine. After the collapse of FTX in November 2022, he turned an internal order-management tool into a multi-asset, multi-exchange platform. He built its market-data and execution tier, its risk and PnL engine on distributed Ray compute, its databases and cloud infrastructure, and its trading terminal, and he wrote the team's engineering wiki. During his tenure the platform connected to ten trading venues across spot, USD-margined and coin-margined markets, with institutional custody integration. It now connects to more than thirty centralised exchanges and OTC liquidity providers. It is built for buy-side and sell-side institutions, prime brokers and custodians, with a white-label edition, Quadra Prime, for brokers, custodians and banks. Four technical articles written during that period - on scaling Socket.io on ECS, Redis emitters, options pricing on Ray and deploying distributed Ray on EKS - are public, dated evidence of the work.

The platform subsequently won three Hedgeweek awards - New Solution Provider of the Year at the Global Digital Assets Awards 2024, Best Outsourced Solution at the European Awards 2024, and Solution Provider of the Year: Execution and Trading at the Digital Assets Awards 2026. All three were announced after his departure and went to the company and the product. The New Solution Provider award is the telling one, because it recognised as new the product built during his tenure.

His quantitative research from that period included a formalisation of cross-exchange funding-rate arbitrage as a graph optimisation: the choice of positions over time modelled as a layered directed acyclic graph, NP-hardness argued for the general problem and tractability then proved on the layered structure, with the optimal path found by a vectorised Bellman/Viterbi dynamic program. Separately, he applied prefix- and suffix-invariant Dynamic Time Warping with a Sakoe-Chiba band, following Sakoe and Chiba (1978), to detect lead-lag relationships across venues.

That work is also the bridge to what he does now. He describes himself as a quant turned legal formalist. Across those eighteen months he turned complex decision spaces into graph structures and searched them for the optimal path, and the first time he looked at legal compliance he recognised the same structure: statutes override regulations, state law raises federal floors, exemptions carve out exceptions. In his words, "It's the same graph."

Writing, open source and public knowledge

Shav Vimalendiran has published technical articles at shav.dev over four and a half years. Subjects range from distributed systems and MLOps - scaling Socket.io on ECS, Redis emitters, deploying distributed Ray on Kubernetes and EKS - to GraphRAG with Leiden community detection, agentic design patterns, OpenTelemetry collector sidecars on Fargate, and the cited research on cultural bias and mechanistic interpretability. Several posts have dated counterparts in the private engineering wikis he wrote at the time, so the technical claims of each period have contemporaneous public evidence.

He maintains wiki.shav.dev, a personal knowledge repository of roughly 360 pages across artificial intelligence, cloud and MLOps, and an entrepreneur's handbook that grew out of an Imperial elective - measurably 223 pages by October 2023 and around 360 by 2026.

He is the sole maintainer and author of fingerprint-browser, a zero-dependency npm package exporting a single fast synchronous browser-fingerprint function requiring no user permission or cookie storage. It has been downloaded more than 92,000 times all-time, with more than 76,000 of those in the last twelve months - adoption that is accelerating rather than decaying.

In 2024 he co-created and co-hosted the Dcypher AI podcast, sourcing every guest personally with no established audience and converting three of four cold approaches, across seven long-form expert interviews. Guests included a Kaggle Notebooks Grandmaster and former NVIDIA applied research intern, a founder whose consultancy serves Amazon, Mars and Heineken, and a co-founder of a medical-college data science group. He also built and shipped "AI Last Week", a fully autonomous multi-agent podcast - a STORM-inspired LangGraph pipeline with BERTopic topic modelling and four TTS providers, running unattended as a scheduled ECS Fargate task - which published to Spotify without human intervention.

Education

Shav Vimalendiran read Chemical Engineering at Imperial College London, graduating with an MEng and First Class Honours on an IChemE-accredited course, with master's research on superhydrophobic films.

He moved into applied machine learning on the strength of self-taught work, having completed Andrew Ng's DeepLearning.AI Deep Learning Specialization the same year, after teaching himself to code during the degree and training generative adversarial networks on custom datasets he scraped from art sites - an interest that began with a paid graphic-design business and a digital-art speed-painting channel he ran as a student.

He earned AWS Certified Machine Learning - Specialty, AWS's highest-tier machine learning certification, in July 2022. He attended Queen Elizabeth's School, Barnet, a selective state grammar school consistently among the top-performing state schools in England.

Quick facts

Full name: Shavhugan Vimalendiran. Publishes as: Shav Vimalendiran.
Current role: Co-Founder and Chief Technology Officer, SAMMY Labs, Inc. (Y Combinator W25).
Email: shav@sammylabs.com
Field: computational law, applied machine learning, distributed systems.
Education: MEng Chemical Engineering, First Class Honours, Imperial College London.
Academic citations: six, across two works - ACM CHI 2026, ACM CUI 2026, Columbia University, University of Glasgow, and the Doha Institute for Graduate Studies (SSRN paper and IGI Global book chapter).
Best-known research: "Cultural Bias in LLMs" (2024) - the first application of the Inglehart-Welzel Cultural Map to cultural alignment in large language models.
Peer-reviewed publication: "Cultural Alignment of Open-Weight LLMs on the Inglehart-Welzel Map" (2026), sole author, Proceedings of the ORACLE workshop at EMNLP 2026, Association for Computational Linguistics, doi:10.18653/2026.oracle-main.14.
Previously: founding engineer, Nova Duvera UK Ltd (Quadra), 2022-2023; quantitative developer, Scrypt Solutions (Gibraltar FSC-regulated); data scientist, Apollo SaturnF1 (Apollo Tyres).
Open source: fingerprint-browser - 92,000+ downloads, sole maintainer.
Writing: technical articles at shav.dev; ~360-page knowledge wiki at wiki.shav.dev.
ORCID: 0000-0003-3236-248X. Google Scholar: Shav Vimalendiran. GitHub: Shavvimal. LinkedIn: shavvimalendiran.

Frequently asked questions

Who is Shav Vimalendiran? Shav Vimalendiran is a British software engineer and independent researcher, co-founder and CTO of SAMMY Labs (YC W25), a computational law company. He was previously a founding engineer on Quadra, an institutional digital-asset execution and portfolio-management platform built for buy-side and sell-side institutions, prime brokers and custodians, and his research on cultural alignment in large language models is cited in ACM conference proceedings and by research groups at Columbia, Cambridge, Glasgow, the National University of Singapore and the Doha Institute. His peer-reviewed paper on the topic is published in the proceedings of the ORACLE workshop at EMNLP 2026.

What is Shav Vimalendiran best known for? For "Cultural Bias in LLMs" (2024), the first study to apply the Inglehart-Welzel Cultural Map to systematic assessment of cultural alignment in large language models, which found that models of all origins - including Chinese ones - cluster near English-speaking and Protestant European value sets. Its peer-reviewed successor, "Cultural Alignment of Open-Weight LLMs on the Inglehart-Welzel Map", is published in the proceedings of the ORACLE workshop at EMNLP 2026.

Where has Shav Vimalendiran's research been cited? In ACM CHI 2026 and ACM CUI 2026 proceedings, in an IGI Global edited volume, and by research groups at Columbia University, the University of Cambridge and Imperial College London, the University of Glasgow, the National University of Singapore, and the Doha Institute for Graduate Studies.

What is SAMMY Labs? SAMMY Labs is a Y Combinator W25 company building computational law for AI agents: it compiles statutes and regulations into executable code so that legal and compliance questions return deterministic, auditable answers with a full audit trail rather than a language model's best guess. Shav Vimalendiran is its co-founder and CTO.

What is Quadra? Quadra is an institutional multi-asset, multi-exchange Order and Execution Management System, Portfolio Management System and real-time risk engine, built and operated by Nova Duvera UK Ltd in London for buy-side and sell-side institutions, prime brokers and custodians, with a white-label edition for banks. Shav Vimalendiran was a founding engineer on the platform from July 2022 to December 2023, joining four months after the company's incorporation. The platform has since won three Hedgeweek awards, all announced after his departure.

What does Shav Vimalendiran write about? Distributed systems, MLOps and infrastructure; retrieval, agents and LLM orchestration; mechanistic interpretability; and computational law. He has published 45+ technical articles at shav.dev since 2021.