Thamirawaran Sathiyalogeswaran AI Researcher Efficient machine learning · Compilers for AI

I research how to make large AI models cheaper to adapt and faster to run.

I am an AI researcher, lead author of a NeurIPS 2026 paper and co-author of a CGO 2027 paper with Prof. Jason Mars (University of Michigan). Alongside research I build systems: the compiler, the tools for AI coding assistants and the deployment system of Jac, an open-source, Python-like programming language.

NeurIPS 2026Lead authorCan one model use many separately trained add-ons at once, without retraining them?LoRASpaceOne composition mechanism for a whole pool of LoRA adapters, with a fixed or a per-input mix.
CGO 2027Co-authorCan program analysis let PyTorch compile a whole model instead of stopping part-way?GraphMendSource rewrites that remove graph breaks. With Prof. Jason Mars (University of Michigan).
EngineeringOpen sourceThe Jac languageA Python-like language with AI calls, data, APIs and deployment built in. I work on its compiler, its tools for AI coding assistants, and its deployment system.

01 Research

Research

Peer-reviewed work on making large models cheaper to adapt and faster to run. Lead-author work comes first.

All research

2026

NeurIPS 2026
Accepted
Lead author

LoRASpace: A Pool-Wide Shared Substrate for Static and Dynamic Multi-LoRA Composition

Using many separately trained LoRA adapters in one model at the same time, without retraining them together, for diffusion models and large language models.

Thamirawaran Sathiyalogeswaran · full author list forthcoming

Main conference · Poster

Details

2027

CGO 2027
Accepted

GraphMend: Code Transformations for Fixing Graph Breaks in PyTorch 2

Source-level program transformations that remove graph breaks, so torch.compile captures one uninterrupted FX graph instead of falling back to Python.

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

02 Engineering

Engineering

The systems I build alongside research, mostly for Jac, an open-source, Python-like programming language: its compiler, tools for AI coding assistants, benchmark grading and deployment.

All engineering

Evaluation Infrastructure

Private Grading for a Coding-Agent Benchmark

If a benchmark publishes its tests, AI coding agents can be tuned to pass exactly those tests. JaseciBench keeps its grading tests private.

2026

GitHub Actions
Benchmark design
Coding agents
Jac

Compilers / Languages

Program Analysis in the Jac Compiler

The compiler’s map of the paths a program can take was wrong for several common statements, so the errors and warnings built on it were wrong too.

Jaseci Labs · Mar 2026

Compiler passes
CFG
Type narrowing
Diagnostics

AI Systems

Tools That Help AI Coding Assistants Write Jac

AI coding assistants know little about Jac, a new language, and in my app-building runs smaller models struggled when shown every tool at once.

Jaseci Labs · Mar to Apr 2026

Model Context Protocol
Coding agents
LLM tooling
Python

Developer Infrastructure

Safe Deploys and Fast Rollback for Jac Applications

Deploys could run the wrong Jac version, fail minutes in after databases were already created, and undo a bad release only by rebuilding it.

Jaseci Labs · 2026

Kubernetes
Release engineering
Blue-green deploys
Jac

03 Overview

How the work fits together

My research makes AI models cheaper to adapt and faster to run. My engineering builds the compiler, the tools for AI coding assistants and the deployment system of the Jac language. GraphMend connects the two: it applies compiler techniques to AI model code. Each number marks where a piece of work applies.

Models · research

LoRASpaceUsing many separately trained model add-ons at once, without retraining. NeurIPS 2026, lead author.
GraphMendRewriting model code so PyTorch can compile the whole model. CGO 2027.

Programs · engineering

Coding-agent benchmarkGrading AI coding agents on Jac tasks with tests they never see.
Tools for AI coding assistantsA server that gives coding assistants Jac documentation and examples.
Compiler analysisHow the Jac compiler tracks the paths a program can take, and the type errors it reports.
Safe deploys and rollbackDeploys that run the pinned Jac version and can switch back to the previous release in seconds.