Hello! 👋
I'm Ido Pinto
AI/ML Research Engineer
I'll be at ICML 2026 in Seoul 🇰🇷. If you're there too, let's connect!btw, I'm open to work :)
View resumeAbout Me
M.Sc. in CS from HUJI, advised by Prof. Guy Katz. Interested in LLMs and their applications to formal methods and program verification.
Through my research I've gotten hands-on with the full ML lifecycle. I built a scalable data pipeline for loop invariant generation (accepted @ ICML 2026!). I care a lot about evals, observability, and reproducibility.
These days I'm into RL, LLM inference optimization, and building reliable and useful AI agents. I like bridging research to production, turning ideas that work in a notebook into systems that actually ship. If anything interests you, reach out!
Education
Master of Science in Computer Science
The Hebrew University of Jerusalem
- Advised by Prof. Guy Katz; GPA 97.7 (Cum Laude)
- First author of "Not All Invariants Are Equal" accepted for poster presentation @ ICML (Seoul, July 2026)
- Delivered oral presentation of thesis research @ the RobustifAI Consortium (Siemens, Belgium, Feb 2026)
Relevant coursework
- Advanced Machine Learning
- Bayesian Machine Learning
- Advanced Natural Language Processing
- Information Theory
- Deep Learning & NLP for Accelerating Science
Bachelor of Science in Computer Science
The Hebrew University of Jerusalem
- GPA 87.2
Relevant coursework
- Probability & Statistics
- Data Structures & Algorithms
- Operating Systems
- Communication Networks
- Object-Oriented Programming
- Machine Learning
- Natural Language Processing
- Image Processing
Publications
Not All Invariants Are Equal: Curating Training Data to Accelerate Program Verification with SLMs
ICML 2026 (Poster)Ido Pinto, Yizhak Yisrael Elboher, Haoze Wu, Nina Narodytska, Guy Katz
Introduces WONDA, a data curation pipeline that refines noisy verifier-generated invariants via AST-based normalization and LLM-driven rewriting. Fine-tuning SLMs on this curated data doubles invariant correctness and verified speedup rates; a 4B model matches GPT-OSS-120B utility and approaches GPT-5.2.
Projects
Verifier-in-the-Loop LLM Agents for Code Repair
Designed an iterative repair agent pairing a reasoning LLM with a formal verifier on 497 unverified Dafny programs from DafnyBench. Found counterexamples alone did not improve overall repair rates, but combining runs with and without them reached 62.4% (+8 points over either approach).
Experience
Teaching Assistant (Grader)
The Hebrew University of Jerusalem
- Evaluated student assignments and exams and handled grade rebuttals across three semesters
- Object-Oriented Programming (Fall 2024)
- Machine Learning (Spring 2025)
- Software Engineering & Communication Networks (Fall 2025)