Dr. Sathiskumar Jothi
Entrepreneur, engineer, inventor and scientist, bridging deep research in materials and engineering with AI: from conventional models to LLMs and agentic AI. Open to research collaboration and consulting.
Beveridge: decisions, not essays
Beveridge is a family of open-weight System One decision models. Give it a document and typed questions (yes/no, one of your options, or a level on your scale) and it returns an answer with a calibrated probability for every option, without generating text, on your own machines. Beveridge-Large-C runs on a laptop CPU; Beveridge-Pro-C brings more knowledge on a GPU; the new Beveridge-Pro-C-Web also picks every step of a browser agent. Released under Apache-2.0, free for commercial use.
Beveridge-Bench accuracy (%)
Held-out test, the same cases for every system.
Full results on bev.sjothi.com · pip install bev-decisions
Beveridge can make mistakes; keep a person in the loop for actions that matter.
Scientist at heart. Engineer by training. Inventor and entrepreneur.
Dr. Sathiskumar Jothi is a scientist, engineer, inventor and entrepreneur with decades of experience spanning engineering, materials science, mathematics and AI/ML research and development, from conventional models to LLMs and agentic AI. He serves as CEO & Director of AI, Data and Engineering at BDST.
With a PhD from Swansea University and a patented AI/ML algorithm delivered for Centrica PLC, he has served as a consultant to international institutions, taught as an Assistant Professor / Lecturer at Swansea University, and collaborated with the UK National Physical Laboratory, Max Planck Institutes and Fraunhofer Institutes. He has delivered AI and data solutions for Fortune 500 clients including Centrica, AstraZeneca, Sky/Comcast, Coles, HCL and Deloitte, reducing costs and optimising operations through predictive deployments.
“A lifelong learner focused on AI-driven innovation, automation, emerging technologies, and cloud solutions.”
- PhD, multiscale modelling of aerospace materials, Swansea University
- Patented AI/ML algorithm for Centrica PLC
- Collaborator: NPL UK, Max Planck, Fraunhofer
- Fortune 500 clients across 10+ industries
- Consultant to international institutions
AI & agentic systems
Conventional ML, deep learning, LLMs, generative and agentic AI, in production at scale.
Engineering & materials
Advanced machining, metal-matrix and natural-fibre composites, Industry 4.0.
Cloud-native scale
Scalable AI architectures on Azure, AWS and GCP with modern data engineering.
Frontier work
Cybersecurity for AI and agentic systems, from model and data protection to secure multi-agent orchestration, plus quantum computing, space applications and energy systems.
Where AI meets atoms
A dual research profile: machine intelligence on one axis, physical engineering and materials on the other.
Artificial intelligence
Machine learning, deep learning, LLMs, generative AI, agentic systems and AI agents for engineering applications.
Space & quantum
Doctoral research on Ariane 5 rocket-engine materials, extending into quantum computing and new computational paradigms.
Materials & manufacturing
WEDM of Al-based composites, Mg hybrid MMCs, natural-fibre polymers, process optimisation, selective laser melting of Inconel 625 and solar-cell materials.
Computation
Multiscale and finite-element modelling from grains to components: hydrogen diffusion and trapping, micromechanics of smart composites, and distributed data processing (UK patent GB2585890).
Citations over time
Google Scholar, cumulative citations by year (2013–2026).
Research areas
Share of the research profile (Google Scholar keywords).
Co-author network
79 co-authors across 31 papers with a DOI (publisher records). Node size = joint papers; names shown for co-authors with 2 or more joint papers; lines join people on the same paper. Hover a node for its name.
Research-topic network
61 publications tagged by title keywords. Node size = papers on the topic; line width = papers combining two topics (thickest: 17 papers on hydrogen embrittlement + multiscale modelling).
Explore the research online
Open-weight decision models
Beveridge-Large-C, Beveridge-Pro-C and Beveridge-Pro-C-Web: benchmarks, install, use cases
Patents, projects and publications
The patent, EU FP7 MultiHy, 62 publications with citations and the open-source release
Experiments in your browser
Real data, recalculated on your device, with the code shown
Hydrogen permeation
Lattice diffusion, trapping and grain-boundary fast paths, checked against exact solutions
Automation lab
How much work a decision model can take over at a target accuracy
From mathematical modelling to cloud deployment
A skill set built across academia, national laboratories and Fortune 500 industry projects.
Areas
- AI, ML, LLMs and agentic AI
- Data science, engineering and analytics
- Cloud architecture: Azure, AWS, GCP
- Engineering, materials and manufacturing research
- Forecasting, recommendation and semantic search
- Cybersecurity for AI and agentic systems
- Quantum computing and frontier technology
Tools and platforms
61 publications · 1,539 citations
Selected works with Google Scholar and publisher citation counts. All publications, the patent and projects · Google Scholar
Most-cited works
| Work | Citations |
|---|---|
| – | |
An investigation of micro-mechanisms in hydrogen induced cracking in nickel-based superalloy 718 | 153 |
Multiscale multiphysics model for hydrogen embrittlement in polycrystalline nickel | 33 |
| – | |
Influence of powder characteristics and additive manufacturing process parameters on the microstructure and mechanical behaviour of Inconel 625 fabricated by selective laser melting | 343 |
Prediction of surface roughness and material removal rate in wire electrical discharge machining on aluminum-based alloys/composites using Taguchi coupled machine learning models | 111 |
Influence of materials and machining parameters on WEDM of Al/AlCoCrFeNiMo0.5 metal-matrix composite | 68 |
Artificial neural network to predict the degraded mechanical properties of metallic materials due to the presence of hydrogen | 67 |
Influence of grain boundary misorientation on hydrogen embrittlement in bi-crystal nickel | 63 |
Reduced graphene oxide wrapped hierarchical TiO₂ nanorod composites for improved charge collection efficiency and carrier lifetime in dye sensitized solar cells | 61 |
Development and characterization of areca fiber reinforced polymer composite | 45 |
Localized microstructural characterization of a dissimilar metal electron beam weld joint from an aerospace component | 30 |
Multi-phase modelling of intergranular hydrogen segregation/trapping for hydrogen embrittlement | – |
Meso-microstructural computational simulation of hydrogen permeation test to calculate intergranular, grain boundary and effective diffusivities | – |
Artificial neural network modeling to evaluate and predict the mechanical strength of duplex stainless steel during casting | – |
Grain boundary and triple junction characteristics analytics of additive manufactured Inconel 625 superalloy using selective laser melting | – |
Inventor
Smart home control system
A control system for intelligent smart-home environments, covering automated environment management.
Distributed data processing using clustering
Scalable distributed data processing through intelligent clustering methods (UK patent GB2585890).
Patented AI/ML algorithm for Centrica PLC
A proprietary AI/ML algorithm deployed in the energy sector.
From machine shops to machine minds
Physical engineering research and the AI frontier, joined by one theme: solving hard problems with rigorous science.
CEO & Director of AI, Data and Engineering, BDST
Leading BDST's mission, "Transforming visions into AI-powered realities": from conventional machine learning to LLMs and agentic AI systems.
Founder, BDST Ltd (Big Data Science & Technology)
Founded BDST as "a co-inventor, not another consultancy": AI, agentic AI, data science and cloud solutions across energy, pharma, healthcare, manufacturing, aerospace, oil & gas, telecom and media.
AI solutions for Fortune 500 companies
Predictive AI deployments for Centrica, AstraZeneca, Sky/Comcast, Coles, HCL and Deloitte; a patented AI/ML algorithm for Centrica PLC; cost reduction and operational optimisation at scale.
Consultant to international institutions
AI, data engineering and scientific rigour for cybersecurity and AI-cyber applications at an international institutional level.
Computational, materials and aerospace researcher
61 publications, talks and reports in computational and multiscale modelling, AI-driven engineering, advanced machining, metal-matrix and natural-fibre composites; collaborations with UK NPL, Max Planck and Fraunhofer Institutes; ASTUTE 2020 advanced manufacturing programme.
Assistant Professor / Lecturer, Swansea University
Lectured and supervised in engineering and materials science: 3D printing, airframe structures, Python, product development and AI, alongside research in computational modelling, multiscale simulation and aerospace materials.
MultiHy project, EU FP7 (CORDIS 263335)
Contributed to the €5.1M project coordinated by Fraunhofer: a multiscale framework for hydrogen embrittlement (DFT, molecular dynamics, kinetic Monte Carlo, finite elements) applied to Ariane 5 Vulcain 2 combustion chambers, automotive steels and offshore wind bearings.
PhD, Swansea University, UK
"Multiscale modelling and experimentation of hydrogen embrittlement in aerospace materials": pulse-plated nickel and nickel-based superalloys in Ariane 5 launcher engines, from combustion-chamber manufacture and welding to embrittlement mechanisms across scales.
Built and launched
BDST — Big Data Science & Technology
The AI, data science and cloud company he founded, "a co-inventor, not another consultancy": end-to-end AI/ML, LLMs, agentic AI, forecasting and scalable cloud solutions for energy, pharma, healthcare, manufacturing, aerospace, oil & gas, telecom and media.
bdst.uk ↗MultiHy — multiscale modelling of hydrogen embrittlement
A €5.1M project coordinated by Fraunhofer (€3.39M EU funding): DFT, molecular dynamics, kinetic Monte Carlo and finite elements, validated on Ariane 5 Vulcain 2 combustion chambers, automotive high-strength steels and offshore wind-turbine bearings.
cordis.europa.eu ↗Worked with: Centrica · AstraZeneca · Sky/Comcast · Coles · HCL · Deloitte · NPL UK · Max Planck · Fraunhofer
Articles and research notes
On AI engineering, agentic systems, cloud platforms and building intelligent software. Follow on Medium
Self-evolving procedural graphs: LLM agents are moving from "memory" to "topology", and where current architectures fail in production
Agents structure procedural knowledge into explicit graphs of (procedure, relation, procedure) triplets, and an offline "LLM refiner" mutates the graph when they fail. The benchmark results are strong, but deterministic offline feedback loops break down in enterprise deployments.
Read the digest
- The offline latency penalty. Runtime environments don't wait for an offline optimisation epoch: if an API contract changes or a service returns a new error code, a static procedural graph cannot repair its topology on the fly.
- Node localisation collapse. Mapping noisy real-world feedback to discrete graph nodes assumes states fit predefined schemas; on edge cases the agent cannot identify its active node and falls back to unconstrained generation or repeats invalid steps.
- Combinatorial graph bloat. Handling thousands of edge cases deterministically makes explicit graphs explode in size, congesting the context window and slowing retrieval.
Failure mode, automated logistics agent. Path: Verify order → Check warehouse API → Process refund. During a peak sale the shipping API throws an unmapped error (ERR_SYNC_TIMEOUT_503); the state cannot be localised, the graph offers no guidance, and the system takes an unmapped path (an unauthorised refund, or a crash) until an engineer runs an offline evolution pass.
Where multi-agent design must go next: dynamic runtime node spawning (transient, conditional subgraphs during execution); hybrid latent–explicit memory (explicit graphs for core logic, latent state for edge cases); real-time local graph repair instead of batch evolution.
Design AI agents for engineering applications, part 1
Core architecture for AI agents that assist design engineers: the reasoning behind component choices, mining enterprise data lakes, and flagging failure points without replacing human insight.
Read on Medium ↗AWS SageMaker vs. AWS Bedrock for generative AI: which to choose?
A practical comparison with step-by-step guides to building RAG applications on each, code examples, guardrails and a head-to-head scorecard.
Read on Medium ↗AI software development vs traditional software development, part 1
Explicit rule-based logic against data-driven probabilistic models, with examples and Python code showing how the development paradigm is shifting.
Read on Medium ↗Domain-specific foundation LLMs: on-premises vs cloud, part 1
Control, compliance, cost and scalability trade-offs of building domain-specific foundation models on-premises or in the cloud.
Read on Medium ↗Let's work together
Open to research collaboration, consulting, AI strategy, and data and engineering engagements, including cybersecurity and institutional projects.