About
Research engineer at the intersection of reinforcement learning and large-scale ML systems. My PhD research is a dual-process, reward-gated architecture for real-time test-time learning under non-stationarity, with provable guarantees. Alongside it I work on LLM pre- and post-training (SFT/RLFT) on TPU, distributed synthetic-data generation, and GPU-accelerated computing.
Education
PhD, Computational and Data Sciences · Aug 2022 – Jul 2027 (expected)
Indian Institute of Science, Bengaluru
Advisor: Prof. Sashikumaar Ganesan · Reinforcement learning, LLM post-training
B.Tech, Mining Engineering · Jul 2015 – Jun 2019
Indian Institute of Technology, Kharagpur
Research
A Dual-Process, Reward-Gated Architecture for Test-Time Learning · 2024 – present
PhD research, IISc — advised by Prof. Sashikumaar Ganesan
- Real-time test-time adaptation of RL policies under non-stationarity: a fast-weight learner adapts continually while a deliberate controller escalates only on environment-reward failures.
- Reward-gated escalation via a drift-robust change detector (CUSUM), with provable gate optimality and dynamic-regret guarantees.
- Generator-free instantiation: a quality-diversity behavioural repertoire with trust-region policy search and verify-before-commit.
Projects
A fuller write-up lives on the projects page.
DSDG: Distributed Synthetic Data Generation · Mar 2026 – present
ZenteiQ AiTech Innovations — collaboration, IndiaAI Mission
- Pre-trained LLMs on TPU using MaxText (JAX).
- Built a scalable, distributed data-generation framework for SFT/RLFT over FastAPI, Kafka and YugabyteDB, with Prometheus/Grafana observability.
- Ran supervised (SFT) and RL (RLFT) fine-tuning; served inference with vLLM.
CAESAR: Combat Aircraft Engagement and Strategic AI Response · Mar 2024 – Mar 2026
Defence R&D Organisation (DRDO) — collaboration
- ML framework for end-to-end emitter detection, identification and multi-sensor data fusion for enemy recognition.
SimInhale: Particle Deposition in Human Lung Airways · Sep 2022 – Mar 2023
Indian Institute of Science
- GPU-accelerated particle deposition; flow simulation with ParMooN (PDE solver), parallelised with OpenMP and CUDA.
Experience
Software Engineer, ARC Document Solutions, Kolkata · Jul 2019 – Jul 2022
- Backend microservices in Node.js/TypeScript for a facility-management platform, with S3 storage and OCR-based deep search.
- Common IAM using OAuth2 and SAML, with Redis for caching.
- CI/CD on AWS using Jenkins and Docker.
Teaching Assistant, CCE, IISc · Aug 2023 – Dec 2023
- Course: Practical AI & MLOps
Teaching Assistant, CCE, IISc · Jan 2023 – May 2023
- Course: Introduction to Computing for AI & ML
Skills
- Programming — Python, C++, TypeScript/Node.js, Haskell
- ML & research — PyTorch, JAX, reinforcement learning, LLM post-training, vLLM
- Systems & tools — Docker, Kubernetes, Kafka, YugabyteDB, AWS/GCP, ClearML
Open source
Mine:
- ExpMan — experiment manager in Rust, on crates.io and PyPI.
- SciREX — scientific ML framework in JAX, with the AiREX Lab and ZenteiQ, on PyPI and GitHub.
Contributions to:
Interests
- Research — reinforcement learning, explainable AI
- Software — Haskell, Emacs, XMonad, Nix
Contact
- Email — lokeshm@iisc.ac.in
- Address — IISc Bengaluru, Karnataka, India 560012
- Languages — English, Hindi, Odia, Telugu