Curriculum vitae

Thamirawaran Sathiyalogeswaran

AI Researcher · Efficient machine learning · Compilers for AI

Profile

My research is on machine-learning methods that run on a single consumer GPU. I am lead author of a NeurIPS 2026 paper on composing many LoRA adapters at once without retraining them, and co-author of a CGO 2027 paper with Prof. Jason Mars (University of Michigan) on rewriting PyTorch model code so the compiler can optimise the whole model. My engineering at Jaseci Labs is on Jac, an open-source, Python-like programming language: its control-flow and type analysis, a Model Context Protocol server that gives AI coding assistants its documentation, and version pinning, rollback and blue-green deploys for Jac applications.

Publications

Thamirawaran Sathiyalogeswaran · full author list forthcoming

Lead author: method, complete codebase and all experiments, on a single RTX 3090. Final-year undergraduate project, University of Moratuwa.

Conference on Neural Information Processing Systems (NeurIPS 2026) · Main conference · Poster

Savini Kashmira, Jayanaka Dantanarayana, Thamirawaran Sathiyalogeswaran, Krisztian Flautner, Lingjia Tang, Jason Mars

Co-author.

International Symposium on Code Generation and Optimization (CGO 2027)

Experience

Jaseci Labs

2024 to 2026

Fully remote, US-based team

AI and Compiler Engineer, full-time

2025 to 2026
  • Maintainer on jaseci-labs/jac with review and merge rights. Merged pull requests are public.
  • Rewrote control-flow graph construction as a stateless entry/exit visitor with short-circuit boolean edges, try / except / finally, match and terminal-statement handling (#5040); union member-access diagnostics (#5010).
  • Model Context Protocol (MCP) server that gives AI coding assistants Jac documentation and examples: knowledge-map and resource-fetching tools, lazy example loading, and a lite / standard / full mode framework; MCP client support in ByLLM (#5362, #5216, #5682, #5474).
  • Version pinning in the Jac stack: a deploy runs the exact Jac release the app pins, for the pod binary, sealed bundle, client bundle and database provisioning, with the pin checked before anything is provisioned (#7713, #8037, #8011, #8777, #8961).
  • Rollback across all services of an app, identified by bundle digest, with a bundle retention contract (#8304, #8120).
  • Blue-green deployments in jacBuilder: the previous version stays as a warm standby, and rollback picks the fastest path (switch to standby, revert revision, or rebuild), tested end to end on a real cluster.
  • Developer documentation and examples (JacBook, LittleX).

AI Systems Research, part-time

Mar to Jun 2025
  • Led a small team fine-tuning and evaluating vision-language models for car-damage captioning: Gemma 3 12B against Qwen 2.5-VL 7B and 3B, LoRA fine-tuning, CIDEr, SPICE and embedding-similarity scoring, with Hydra configs and Weights & Biases tracking.
  • Key-information extraction from scanned receipts: LoRA on Gemma 3 12B, deterministic inference, field-level Levenshtein scoring.

Compiler Engineer, internship

Sep 2024 to Mar 2025
  • Parser, grammar and code-generation fixes; compiler caching and incremental-build experiments.

Education

BSc Engineering (Hons), Electronic and Telecommunication Engineering

2022 to 2026

University of Moratuwa. Coursework completed June 2026; graduating 11 December 2026.

  • Final-year project: multi-LoRA composition (LoRASpace, NeurIPS 2026).
  • Linear Algebra A+, Calculus A.

Awards

  • Mahapola Higher Education Merit Scholarship (Sri Lankan national merit scholarship).
  • GCE Advanced Level 2020: ranked 34th in Sri Lanka (Z-score 2.7488).
  • Second place, All Island Olympiad Mathematics Quiz, 2019.

Technical areas

ML research
PyTorch, Hugging Face Transformers and Diffusers, PEFT (LoRA, QLoRA), LLM-as-judge evaluation, single-GPU experiment design
Compilers
Parsing, AST and IR, control-flow and data-flow analysis, type narrowing, code generation, PyTorch 2 graph capture
Agents & evaluation
MCP server and client design, hidden-test benchmark design, agent repair loops
Systems
Kubernetes, Docker, graph persistence over MongoDB and SQLite, GitHub Actions, Linux
Languages
Python, C++, C, TypeScript, Jac, SQL