Building responsible, production-grade AI systems — from foundational research to deployed models.
About
I am a data scientist and AI researcher with a PhD in Data Science from University College London. My work sits at the intersection of large language models, responsible AI, and applied machine learning — turning cutting-edge research into systems that are reliable, interpretable, and useful in the real world.
I started in aerospace engineering, where I learned to reason about complex systems under uncertainty — modelling dynamics, propagating error, and designing for safety. That mindset now drives my AI research: I care not just about whether a model works, but about whether it can be trusted, explained, and relied upon.
Today I lead research on trustworthy large language models, with a focus on calibrated uncertainty, alignment, and fairness evaluation. I work across the full stack — from mathematical foundations to production deployment.
University College London (UCL)
Thesis: "Calibrated Uncertainty and Alignment in Large Language Models." Research focused on trustworthy LLMs — uncertainty quantification, alignment, and evaluation under distribution shift.
University College London (UCL)
Distinction. Specialised in statistical machine learning, probabilistic modelling, and deep learning. Dissertation on variational inference for sequence models.
SRM University
Foundations in applied mathematics, fluid dynamics, control systems, and computational modelling — the analytical backbone for a move into data science.
Skills & Expertise
From foundational research to production MLOps — the capabilities I use to take an idea from hypothesis to a reliable, monitored system.
Modelling, LLMs and deep architectures
Shipping models into reliable systems
Pipelines, warehouses and vector search
Interpretability, fairness and safety
Tools & frameworks
Research Papers
Five recent papers on the topics defining modern AI — trustworthy LLMs, alignment, efficient inference, and fairness evaluation.
S. Ravichandran, A. Mehta, L. Okafor
We introduce a distribution-free conformal prediction framework that produces statistically valid prediction sets for autoregressive LLMs, reducing overconfident hallucinations by 38% while preserving answer utility across eight benchmark datasets.
S. Ravichandran, M. Chen, P. Verma
A robust variant of DPO that maintains alignment guarantees when the preference data is drawn from a shifted distribution. We prove a generalisation bound and demonstrate stable reward modelling under covariate shift on three open-source base models.
S. Ravichandran, T. Albrecht, R. Iyer
We present MoE-Edge, a conditional-computation scheme that activates only 4% of parameters per token while matching dense 7B-model quality. Deployed on commodity hardware, it achieves 3.1× throughput with under 2% quality degradation.
S. Ravichandran, J. Park, N. Costa
A disentangled representation framework that recovers latent causal variables from high-dimensional temporal signals using interventional data. We show identifiability guarantees and outperform baselines on climate and healthcare benchmarks.
S. Ravichandran, K. Osei, L. Okafor, M. Chen
FaIR-Eval defines 12 measurable fairness traits across language, vision, and multimodal foundation models. We audit 14 commercial APIs, surface systematic gaps, and release an open toolkit adopted by three national AI safety institutes.
Research philosophy: I publish open-access wherever possible, release code and datasets alongside papers, and write for practitioners as well as academics. Science only matters if someone can use it.
Latest Publications
Journal articles, a book chapter, and invited reviews from the past two years — spanning agentic systems, RAG at scale, healthcare ML, and the bridge from engineering to AI.
Journal of Machine Learning Research (JMLR) · 2025
A field study on deploying tool-using LLM agents in enterprise settings — covering planning reliability, tool-call verification, and a novel self-correction loop that reduces failed tool invocations by 44%.
View publicationSpringer — Foundations of Modern AI · 2025
A comprehensive chapter on building production RAG systems: chunking strategies, hybrid retrieval, re-ranking, evaluation harnesses, and cost-aware serving patterns for billion-document corpora.
View publicationNature Machine Intelligence · 2024
We combine deep survival analysis with SHAP-based interpretability to forecast sepsis onset 6 hours before clinical deterioration, validated on a multi-site cohort of 240,000 ICU stays.
View publicationNeurIPS — Datasets & Benchmarks · 2024
A diffusion-based synthetic data generator with formal differential-privacy guarantees. We show downstream models trained on synthetic data match real-data performance within 3% while bounding privacy loss.
View publicationRoyal Society Interface · 2023
An invited review tracing how methods from aerospace engineering — reduced-order modelling, uncertainty propagation, and control theory — inform modern machine-learning system design.
View publicationExperience
Walmart
Research and deploy large-scale AI models across retail and supply-chain use cases. Build production ML systems spanning forecasting, personalization, and responsible AI, serving millions of customers and associates.
DeepMind Applied
Managed a team of six on trustworthy LLM research and deployment. Owned the evaluation and safety tooling for a flagship assistant product; shipped calibrated-uncertainty inference used by 4M+ monthly users.
Turing Institute
Research scientist on the AI safety programme. Published on alignment, fairness auditing, and uncertainty; co-authored the institute's responsible-AI deployment framework adopted by two government agencies.
Independent
Advised fintech and healthcare startups on ML strategy, MLOps, and model risk. Delivered production forecasting and NLP systems serving over 50M requests per month.
Get in touch
Open to research collaborations, advisory roles, and speaking engagements. Drop me a note and I'll get back to you within a few days.