Computing’s Top 30: Sachin Kumar

Sachin Kumar is one of our "Computing's Top 30 Early Career Professionals" for 2025. This program seeks to highlight an esteemed group of rising stars who earned this honor for their exceptional early-career achievements and role in driving advancements across the computing landscape. 

Introduction

My name is Sachin Kumar, and I am a Senior Data Scientist III at LexisNexis in Raleigh, North Carolina. Over the past decade, I have dedicated my career to building AI systems that solve meaningful, real-world problems, starting from early work in natural language processing and search, to today leading the development of AI agents that will help attorneys streamline and automate their workflow more efficiently. My work sits at the intersection of cutting-edge research and production-grade engineering: I don't just prototype ideas, I ship them at scale. Most recently, I have also been deeply involved in AI safety and alignment research as a Research Mentee with the Supervised Program for Alignment Research (SPAR), where I worked to develop evaluation benchmarks that measure the risk of AI agents causing irreversible societal harm.

What do you consider your highest achievement so far and how do you plan to continue or build on that success?

I consider my AI safety research specifically my paper 'Overriding Safety Protections of Open-Source Models' — one of my most meaningful achievements. Through systematic experimentation, I demonstrated that harmful fine-tuning can meaningfully erode the safety protections of open-source models, while safety-focused training can substantially restore and strengthen them. This surfaced a real and underappreciated vulnerability in how open-source models are deployed and gave the research community concrete, empirical evidence to act on. Building on this, my SPAR project on evaluating 'lock-in' risks in autonomous LLM agents pushes these concerns further — designing benchmarks that reveal behavioral divergences between what agents claim they will do and what they actually execute. My goal is to release this as an open-source tool and embed it into pre-deployment safety pipelines. If I can help make safety evaluation a standard part of AI development rather than an afterthought, that would be a deeply meaningful contribution.

How are you currently involved in the tech community aside from your job (volunteering, open-source projects, mentoring, etc)?

Staying connected to the broader research community is something I invest in deliberately. I contribute to open-source projects in the AI safety and NLP space, and I am actively working on several research efforts across AI interpretability, safety, and alignment, with conference submissions in progress. Giving back to the peer review process is also important to me , as I have served as a reviewer for AI and ML conferences, evaluating research submissions and helping maintain the quality of work that enters the field.

I also write a Medium blog where I break down and analyze the latest AI research papers, translating dense technical findings into accessible insights for practitioners and enthusiasts alike. I find that the discipline of explaining a paper clearly forces a deeper understanding of it — and if it helps others stay current with a field that moves as fast as ours does, all the better.

Is there any emerging technology or industry segment you find exciting or interesting?

Agentic AI and autonomous systems capable of multi-step reasoning and action, is both the most exciting and most consequential development happening right now. The shift from AI as something you query to AI as something that acts on your behalf introduces fundamentally new capabilities and fundamentally new risks. I have built and deployed multi-agent systems at Chegg for personalized education and am now applying similar approaches to the legal domain at LexisNexis, where accuracy and reliability are non-negotiable. That hands-on experience is precisely what draws me to the safety side of this work, the more powerful these agents become, the more critical it is that we can verify their behavior before deployment

How do you see technology shaping humanitarian efforts or social good in the next 5 years?

I believe AI's most transformative humanitarian impact in the next five years will come through two channels: democratizing access to expert knowledge, and accelerating disaster response.

My AAAI Fall Symposium 2025's paper publication titled 'Evolve-DGN: An Evolving Dynamic Graph Network for Adaptive and Equitable Resource Allocation in Disaster Response' reflects my personal investment of research efforts in this area. Disaster response has historically been constrained by the speed of human coordination and the inequitable distribution of resources. Graph-based AI systems that adapt in real time to evolving crisis conditions can help ensure that aid reaches the most vulnerable populations more efficiently and equitably.

Beyond disaster response, AI agents that can reason over legal, medical, and scientific knowledge have the potential to extend access to expertise that was previously gated by geography and socioeconomic status. The attorney workflows I am working to streamline at LexisNexis is one small example , the same paradigm can be applied to healthcare or education in underserved communities, could be transformative. The critical challenge is ensuring these systems remain safe, fair, and accountable.

What advice would you give to young professionals or recent graduates who are trying to enter your field?

My first piece of advice is simple: build real things, and then care deeply about why they work. Early in your career, the temptation is to collect tools and frameworks. What actually compounds over time is the discipline of shipping something, measuring its impact honestly, and developing the intellectual curiosity to understand the system beneath the surface. That habit — more than any certification or course — is what makes you adaptable as the field evolves.

Second, if you are entering AI today, take safety and alignment seriously. We are at a genuinely pivotal moment. The researchers who will matter most in the next decade are not just those who can build capable systems, but those who understand how to make them trustworthy and robust. Engage with interpretability, reward modeling, and alignment literature even if your current role doesn't require it — that depth will set you apart.

Finally, seek out cross-disciplinary experiences wherever you can. My background spans working on different domains like IVR systems, legal research, computational advertising, ed-tech, and AI safety and alignment research. Each one forced me to think about problems differently and communicate technical ideas to non-technical stakeholders. That breadth doesn't dilute your expertise; it makes it more powerful.

Stay Connected

Find Sachin Kumar on LinkedIn.

Make Your Voice Heard

Build connections in the field and contribute your expertise by submitting your research to one of our Calls for Papers.