Head of Generative AI, KGiSL Centre for Innovation
Mathi Yuvarajan T.K.
Engineer · Software Developer · AI Researcher · Generative AI Leader
I build AI systems that survive contact with production, and I help engineers around me do the same.
Profile
Engineer. Researcher. Educator. Builder.
Head of GenAI · KGiSL
Coimbatore, Tamil Nadu, India
I lead Generative AI at KGiSL Centre for Innovation. Most days that means three things: shipping AI features people actually use, keeping the research pipeline honest, and making sure the engineers on my team can do both without me in the room.
I started in embedded systems, writing firmware where a wrong register costs you a week. That taught me to respect latency, memory, and failure modes. Web and mobile taught me product. DevOps taught me that nothing counts until it deploys. AI is where all three finally converge.
The other half of my time goes to teaching. I train engineers, mentor student teams, and run community sessions. It is the highest leverage thing I know how to do.
The Journey
Software Engineering
Systems, C/C++, and the habit of reading code before writing it
Embedded Systems
Bare-metal firmware on STM32 and ARM Cortex, protocol work close to the wire
Mobile & Web Development
Full-stack product engineering, front to back
DevOps
CI/CD, containers, and infrastructure that does not page you at 2am
Artificial Intelligence
Machine learning, neural networks, computer vision
Generative AI
LLMs, RAG, vision-language models, and the eval work behind them
Agentic Systems
Multi-agent orchestration, MCP, tool-using systems with real guardrails
Research & Innovation
Leading applied AI research and innovation strategy
Experience
Where I've built and led.
Firmware, then products, then applied AI research. Each step made the next one make sense.
Nov 2024 · Present
CurrentHead of Generative AI
KGiSL Centre for Innovation
I own the Generative AI direction here: what we research, what we ship, and who builds it. That covers agentic systems in production, the eval work that keeps them honest, and mentoring the teams doing the work.
Jul 2024 · Apr 2025
Embedded Software Engineer
Robert Bosch
Bare-metal firmware for automotive-grade systems. Low-level communication protocols, safety-critical code, and the discipline that comes with software you cannot patch over the air.
Jun 2022 · Jul 2024
Product Engineer
Codingmart Technologies
Shipped full-stack products end to end, from architecture and backend services through to web and mobile. Learned what actually breaks once real users show up.
Education
Master of Technology, Artificial Intelligence and Machine Learning
Birla Institute of Technology and Science, Pilani
Postgraduate research in AI/ML systems and applications
Bachelor of Engineering, Electrical and Electronics Engineering
Anna University
Foundational engineering education across electrical, electronics, and systems design
Research
Following the frontier of language model research.
A running reading list across pretraining, reasoning, alignment, and agentic systems. These papers shape how I think about applied Generative AI.
Pretraining
- Parity-Aware Byte-Pair Encoding: Improving Cross-lingual Fairness in Tokenization
- Chameleon: A Flexible Data-mixing Framework for Language Model Pretraining and Finetuning
Embeddings
- MMTEB: Massive Multilingual Text Embedding Benchmark
- Improving Text Embeddings with Large Language Models
- DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model (Section 2.1)
- NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models
Mixture-of-Experts
- Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer
- Your Mixture-of-Experts LLM Is Secretly an Embedding Model for Free
Reasoning & Alignment
- BIRD: A Trustworthy Bayesian Inference Framework for Large Language Models
- Direct Preference Optimization: Your Language Model is Secretly a Reward Model
- Training a Generally Curious Agent
Agent Harness
- Agent Workflow Memory
- The OpenHands Software Agent SDK: A Composable and Extensible Foundation for Production Agents
Long-Context & Latent Attention
- Efficient Streaming Language Models with Attention Sinks
- TransMLA: Multi-Head Latent Attention Is All You Need
- DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models (Section 2)
Reinforcement Learning
- DAPO: An Open-Source LLM Reinforcement Learning System at Scale
- Understanding R1-Zero-Like Training
- Optimas: Optimizing Compound AI Systems with Globally Aligned Local Rewards
- CWM: An Open-Weights LLM for Research on Code Generation with World Models
Self-Play & Test-Time Scaling
- Absolute Zero: Reinforced Self-play Reasoning with Zero Data
- SPICE: Self-Play In Corpus Environments Improves Reasoning
- Toward Training Superintelligent Software Agents through Self-Play SWE-RL
- Scaling LLM Test-Time Compute Optimally Can be More Effective than Scaling Parameters for Reasoning
- Learning to Discover at Test Time
Mode Collapse & Linear Transformers
- Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)
- Linear Transformers Are Secretly Fast Weight Programmers
- Parallelizing Linear Transformers with the Delta Rule over Sequence Length
Diffusion Language Models
- Diffusion-LM Improves Controllable Text Generation
- Large Language Diffusion Models
Safety & Interpretability
- On the Role of Attention Heads in Large Language Model Safety
- Emergent Misalignment: Narrow Finetuning Can Produce Broadly Misaligned LLMs
- Scaling and Evaluating Sparse Autoencoders
- Sparse Crosscoders for Cross-Layer Features and Model Diffing
Calibration & Scaling Laws
- Active Task Disambiguation with LLMs
- Learning to Route LLMs with Confidence Tokens
- Training Compute-Optimal Large Language Models
- Scaling Laws for Precision
How These Systems Work
No magic, just weights and plumbing.
The mental model I work from, and the one I teach. If you can picture the forward pass, you can reason about cost, latency, and failure.
Self-Attention
Every token scores every token before it. The mask is what makes generation causal, and the scores are where most debugging actually happens.
Forward Pass
Embed, attend, normalize, project, sample. Five steps repeated a few dozen times. Knowing where latency and cost accumulate is most of the job.
Learned Weights
Under the abstraction it is still a network of weighted sums. I keep that model in mind whenever an output looks like magic or like nonsense.
Selected Work
Systems and products I've built.
Agentic AI, LLMOps, MLOps, and the systems work underneath. Most of these started as a problem someone had at 2am.
Agentic AI
Agent Orchestration Runtime
An MCP-based runtime where planner, retriever, and executor agents share one tool registry. Every tool call is typed, budgeted, and replayable, so a failed run can be traced step by step instead of guessed at.
Agentic AI
Self-Correcting RAG Pipeline
Retrieval with a critic loop. The model grades its own grounding, reruns retrieval when confidence drops, and refuses to answer when the context is thin. Cut unsupported answers by a wide margin on our internal eval set.
LLMOps
LLM Evaluation Harness
A regression suite for prompts. Golden datasets, LLM-as-judge scoring with human spot checks, and a CI gate that blocks a prompt or model change if quality drops. Prompts get treated like code because they behave like code.
LLMOps
Prompt & Model Registry
Versioned prompts, model configs, and system instructions behind a single API. Rollouts are staged, every response carries the version that produced it, and rollback is one flag rather than a redeploy.
LLMOps
Inference Gateway & Cost Control
One gateway in front of several model providers. Semantic caching, token budgets per team, streaming passthrough, and automatic fallback when a provider degrades. Cost per request became a dashboard number instead of a monthly surprise.
MLOps
Training to Serving Pipeline
End to end path from feature store to endpoint. Experiment tracking, model registry, containerized serving, and canary rollouts on Kubernetes. Retraining is a scheduled job, not a person remembering to run a notebook.
MLOps
Drift & Quality Monitoring
Production monitoring for models that quietly go stale. Input drift, prediction drift, and slice-level accuracy tracked against a baseline, with alerts that name the affected segment instead of firing a generic threshold.
Systems
Vision Language Inspection System
A VLM pipeline that reads images and structured context together to flag defects, with a small distilled model on the edge and the large model reserved for the uncertain cases.
Systems
Embedded IoT Monitoring System
Bare-metal firmware and communication stack for a real-time monitoring device on STM32, speaking CAN and LIN with a hard latency budget and no room for a garbage collector.
Technical Expertise
Tools and technologies I build with.
Bare-metal firmware at one end, large language model systems at the other. The range comes from actually walking that path.
Artificial Intelligence
AI Ecosystem
Software Engineering
Cloud & DevOps
Embedded Systems
Tech Community Building
Knowledge compounds when it's shared.
I run two of the largest tech communities in Coimbatore. Both exist for the same reason: engineers learn faster in a room with other engineers.
GenAI Coimbatore
Founder & Lead
One of the largest Generative AI communities in the region. Hands-on sessions on LLMs, RAG, agents, and evals, run for people who want to build rather than watch slides.
IPS Tech Community
Chief Technology Officer
I set the technical direction: what we teach, how we run build events, and how we keep the bar high as the community grows. The goal is engineers who ship, not attendees who collect certificates.
Robotics · Hands-on
OpenClaw build session
- Running technical workshops on AI and software engineering
- Training students and early-career developers
- Mentoring developers and student innovation teams
- Organizing hackathons and innovation events
- Leading open-source initiatives
- Growing AI and developer communities
- Running technical bootcamps
200+
Innovation Projects Evaluated
1000+
Students & Developers Reached
2
Tech Communities Led
Training & Mentoring
Learn. Build. Experiment. Publish. Innovate.
My teaching loop. Learning stays theoretical until someone builds with it, so I optimize for the shortest path to a working thing.
Learn
Build
Experiment
Publish
Innovate
Topics I Teach
Beyond Technology
What keeps me curious.
The things I do when nobody is paying me to. They shape the work more than the job title does.
Exploring New Technologies
Research
Building Side Projects
Reading Technical Papers
Teaching
Community Building
Product Ideation
Electronics & Hardware Experimentation
“கற்றதனால் ஆய பயனென்கொல் வாலறிவன் நற்றாள் தொழாஅர் எனின்?”
Knowledge is only worth what it becomes once someone else uses it.
Engineer by Profession | Teacher by heart ❣️ | அறம் செய்ய விரும்பு
Technology gets interesting at the point where learning turns into a working prototype, and the prototype turns into something people rely on.
Contact
Let's build something interesting.
An AI system, a product idea, a research collaboration, or a workshop for your team. Tell me what you are working on.