AI-Integrated Programming · Languages · Runtimes

Jayanaka Dantanarayana

I build programming languages, compilers, and runtimes that make AI-Integrated software simpler to write and more reliable to run. I am a PhD student in Computer Science and Engineering at the University of Michigan, working with the Jaseci Research Lab, and the creator of Sigil, the skill compiler.

Sigil ↗ GitHub Google Scholar DBLP LinkedIn

Latest work: Sigil

The skill compiler

Don't prompt the skill. Compile it.

Agent skills today are prose: a model reads a SKILL.md, re-derives its control flow on every run, and may silently skip mandated steps. Sigil takes the opposite approach. It compiles the Markdown skill into a typed agent harness, so every step, rule, and check is enforced by program structure rather than hoped for in a prompt.

Sigil lifts the skill into AG-IR, a typed agentic intermediate representation that separates model-owned cognition from code-owned mechanism, audits it with a gate battery, and mechanically lowers it to a runnable artifact. It is built in Jac on the Jaseci stack, works with Claude Code without an API key, supports 52 providers plus local models, and ejects a self-contained single-file executable.

86%

mandated steps executed, vs 56% as a prompt

2.3×

full-procedure completion

0.58×

tokens vs the prose baseline

sigilagent.com · GitHub · Docs · Paper

# one command: the compiler and the whole agent
$ curl -fsSL https://github.com/sigilagent/sigil/releases/latest/download/install.sh | bash

# SKILL.md in, a typed agent harness out
$ sigil compile ./SKILL.md -e agent.jac
✓ LIFT   SKILL.md → AG-IR
✓ GATE   coverage · drift · structure
✓ LOWER  AG-IR → agent.jac

$ ./agent.jac "extract the tables from report.pdf"

Research

Software development is shifting from writing conventional code to composing programs with LLMs inside them. My research asks what languages, compilers, and runtimes should look like in that world, so that developers express intent once in code and systems carry it the rest of the way.

Meaning-Typed Programming

MTP replaces prompt engineering with a single language-level abstraction: the by operator invokes an LLM through the meaning already present in code, backed by MT-IR at the compiler level and MT-Runtime at execution. Published at OOPSLA 2025 and shipped as byLLM in the Jac language.

Semantic Engineering

Where code alone is not enough, SemTexts let developers attach lightweight natural-language context directly to program constructs. Additional intent lives in the program, not in handcrafted prompts.

Agent Skills as Programs

Sigil treats an agent skill as source code: prose procedures are lifted into AG-IR, a typed agentic IR, and lowered into harnesses that enforce the procedure by construction rather than by instruction.

Graphs for Retrieval and ML Systems

TOBUGraph and GraphRunner move retrieval beyond chunk-similarity RAG with knowledge-graph structure, deployed in a production memory application. GraphMend repairs graph breaks in PyTorch 2 through source-level code transformation.

Publications

SIGIL: Compiling Agent Skills into Typed Harnesses

Jayanaka L. Dantanarayana, Savini Kashmira, Lingjia Tang, Jason Mars

arXiv preprint · 2026

arXiv Project Code
Prompt Less, Smile More: MTP with Semantic Engineering in Lieu of Prompt Engineering

Jayanaka L. Dantanarayana et al.

arXiv preprint · 2025

arXiv
MTP: A Meaning-Typed Language Abstraction for AI-Integrated Programming

Jayanaka L. Dantanarayana, Yiping Kang, Kugesan Sivasothynathan, Christopher Clarke, Baichuan Li, Savini Kashmira, Krisztián Flautner, Lingjia Tang, Jason Mars

OOPSLA 2025 · Proc. ACM Program. Lang. 9 (OOPSLA2)

DOI arXiv Artifact
GraphMend: Code Transformations for Fixing Graph Breaks in PyTorch 2

Savini Kashmira, Jayanaka L. Dantanarayana, Thamirawaran Sathiyalogeswaran, Krisztián Flautner, Lingjia Tang, Jason Mars

arXiv preprint · 2025

arXiv
GraphRunner: A Multi-Stage Framework for Efficient and Accurate Graph-Based Retrieval

Savini Kashmira, Jayanaka L. Dantanarayana, Krisztián Flautner, Lingjia Tang, Jason Mars

arXiv preprint · 2025

arXiv
TOBUGraph: Knowledge Graph-Based Retrieval for Enhanced LLM Performance Beyond RAG

Savini Kashmira, Jayanaka L. Dantanarayana, et al.

EMNLP 2025 · Industry Track

arXiv
Scaling Down to Scale Up: A Cost-Benefit Analysis of Replacing OpenAI's LLM with Open Source SLMs in Production

Chandra Irugalbandara, et al., incl. Jayanaka L. Dantanarayana

IEEE ISPASS 2024

arXiv
Surface Texture Reproduction and Amplification for Haptic Perception

Jayanaka L. Dantanarayana, et al.

IEEE IECON 2023
Haptic Based Surface Texture Classification Using Machine Learning Techniques

Savini Kashmira, Jayanaka L. Dantanarayana, et al.

IEEE IES OnCon 2023

Full list on Google Scholar and DBLP.

About

I started in electrical engineering at the University of Moratuwa, Sri Lanka, where I worked on haptics and surface-texture perception. That systems background carried me to the University of Michigan, where I am advised by Jason Mars and Lingjia Tang in the Jaseci Research Lab, working at the intersection of programming languages, compilers, and AI systems.

My work ships in the open source Jac programming language and the Jaseci runtime stack: MTP is available as the byLLM plugin, and Sigil is built end-to-end on the same foundation. I presented MTP at SPLASH/OOPSLA 2025 in Perth, and I am co-organizing the Jaseci tutorial on agentic AI at ASPLOS 2026.