Abel Yagubyan

Senior Data Scientist at C3.ai

Hello! I'm a Senior Data Scientist at C3.ai with expertise in Deep Learning, Predictive Maintenance, and High-Performance Computing. I hold an M.S. in Computer Science from Northwestern University (Summa Cum Laude) and dual B.A. degrees in Computer Science & Applied Mathematics from UC Berkeley.

Previously, I co-founded FibonAI (UC Berkeley Skydeck), conducted research at Lawrence Berkeley National Laboratory on UPC++ performance testing, and interned at Apple. I've also reviewed 100+ papers for Elsevier's AI and Vision Computing journals.

191

Peer Reviews

Citations

Merged PRs

Journals

Citing Groups

News

Named Area Triager on TruLens (the open-source LLM and agent evaluation library maintained by Snowflake) and added to MAINTAINERS.md.

Sep 2026

Contribution credited in inspect_evals v0.20.0, the UK AI Security Institute's evaluation suite for Inspect AI.

Sep 2026

"The Coin Flip Judge?" accepted at GroundLM 2026 (EMNLP 2026) as an Archival Long Paper; camera-ready submitted.

Sep 2026

Citations

Work cited by research groups at Meta, Alibaba and Nanjing University, USTC, University of New Mexico, and KTH.

Aug to Sep 2026

Preprints

Released two sole-author preprints on LLM agent reproducibility and LLM-as-a-judge reliability.

Apr 2026

Promoted to Senior Data Scientist at C3.ai in under 18 months for high-impact customer solutions

2024

Co-founded FibonAI, accepted into UC Berkeley's Skydeck Pad-13 Incubator competing against 5,000+ startups

June - Dec 2023

Graduated with M.S. in Computer Science from Northwestern University with Summa Cum Laude Honors

June 2023

Research contributor at Lawrence Berkeley National Lab's Pagoda Project on UPC++ performance testing

June 2022 - April 2023

Published "Embedding of Programming IDEs into Computer-Based Testing Software" at ACM SIGCSE '22

March 2022

Graduated from UC Berkeley with dual B.A. in Computer Science & Applied Mathematics

May 2022

Software Engineering Intern at Apple x UC Berkeley, managing CS61C projects for 4000+ HBCU students

May - Aug 2021

Research & Publications

The Coin Flip Judge: conceptual illustration of an unreliable coin-flip verdict on an unstable balance scale

GroundLM 2026 (EMNLP), Archival

The Coin Flip Judge? Reliability and Bias in LLM-as-a-Judge Evaluation

Abel Yagubyan (sole author)

arXiv:2606.13685, April 2026. Accepted as an Archival Long Paper at the GroundLM 2026 workshop at EMNLP 2026, Budapest (proceedings forthcoming).

TLDR: Repeated identical LLM-as-a-judge evaluations across six judges from three providers show pairwise verdicts flip on 11 to 14 percent of re-runs, order bias persists across model generations, a single trial recovers only 86.6 percent of the consensus verdict, and about 11 trials are needed to stabilize a verdict; argues for multi-trial aggregation, order randomization, and uncertainty reporting as standard practice.

arXiv abstract PDF

LLM agent reproducibility: conceptual illustration of one task branching into two divergent tool-call pipelines

Preprint

How Consistent Are LLM Agents? Measuring Behavioral Reproducibility in Multi-Step Tool-Calling Pipelines

Abel Yagubyan (sole author)

arXiv:2605.28840, April 2026.

TLDR: Measures whether an LLM agent given the same task twice takes the same actions across several models and providers, and documents that agents with large unconstrained action spaces vary run to run.

arXiv abstract PDF

Embedding of Programming IDEs into Computer-Based Testing Software

Abel Yagubyan, Dan Garcia

ACM SIGCSE '22, Technical Symposium on Computer Science Education

TLDR: Published peer-reviewed paper presenting an interactive RISC-V compiler integrated into PrairieLearn for online programming education, enabling 3,000+ students to complete CS61C exams with embedded IDE functionality.

Paper ACM DL

SCALPEL: Customized Deep Neural Network Compression

Abel Yagubyan

Northwestern University, Research Project (Jan 2022 to June 2022)

TLDR: Developed Python-based DNN compression framework that reduced model sizes (AlexNet, LeNet-5) by up to 80% and achieved 3.5x speedup by customizing pruning techniques to target hardware parallelism (CPU, GPU, Microcontroller).

Report GitHub

UPC++ Performance Regression Testing & Benchmarking

Abel Yagubyan, Lawrence Berkeley National Laboratory

Pagoda Project, Research Contribution

TLDR: Developed automated performance regression testing platform for UPC++, a C++ library for distributed-memory parallel computing. Implemented distributed hashing benchmarks and compared performance against MPI and SHMEM.

LBNL Pagoda

Citations & Adoption

Cited and built on by research groups at Meta Superintelligence Labs (OmnilingualGAIA2, arXiv:2608.08775), Alibaba Group and Nanjing University (OpenCodeReview, arXiv:2608.09290, open source at github.com/alibaba/open-code-review), USTC (arXiv:2607.02873), University of New Mexico, and KTH Royal Institute of Technology (arXiv:2609.06147). 25 citations total as of September 2026.

Waxell published operational guidance for its users based on the judge-reliability findings, citing the paper as its first source.

Industry Experience

C3.ai

Senior Data Scientist

May 2024 - Present | Redwood City, CA

Lead data science initiatives in predictive maintenance and generative AI across large-scale deployments for Shell, ExxonMobil, Dow, Bloom Energy. Directed pilots delivering over $10M in value and deploying 1,000+ models per project. Developed foundational healthcare transformer model trained on 1M+ U.S. patient records. Created SpeedyREL, an internal tool reducing deployment timelines by 50%, adopted by 5+ projects. Promoted to Senior in under 18 months.

Python PyTorch Transformers XGBoost SQL Kubernetes

FibonAI

Co-Founder

June 2023 - December 2023 | San Francisco, CA

Designed and built SaaS platform providing comprehensive LLM-powered workspace for in-house corporate legal teams—integrating intake & triage, contract management, matter management, and AI Assistants. Successfully onboarded 8 General Counsels/Chief Legal Officers. Achieved 60% conversion rate during demos and 100% weekly user growth. Accepted into UC Berkeley's Skydeck Pad-13 Incubator (competing against 5,000+ startups globally).

Python Flask AWS MongoDB GPT-4 React

Apple x UC Berkeley

Software Engineering Intern

May 2021 - August 2021 | Remote

Project lead in collaboration with Apple and UC Berkeley's CS61C: Computer Architecture course. Coded and managed Computer Architecture, Algorithm, and ML-based projects for 4,000+ students simultaneously. Developed implementations using C, C++, Python, SIMD, OpenMP, and RISC-V Assembly.

C C++ Python RISC-V Assembly OpenMP

Software & Projects

FibonAIProduction

Premier legal AI-powered workspace transforming in-house legal operations with GPT-powered tools.

Private Repo Python 10+ Customers

Website

SCALPELResearch

DNN model compression achieving 80% size reduction and 3.5x speedup through hardware-aware pruning.

Python CUDA

GitHub

UPC++ Benchmarking SuiteFramework

Automated performance regression testing for UPC++ with comparison to MPI and SHMEM.

C++ Python

GitHub

PrairieLearn RISC-V CompilerTool

Interactive RISC-V compiler for online education serving 3,000+ students. Published at SIGCSE '22.

Python 3,000+ Students

GitHub

CS61ClassifyEducational

Neural network implementation in RISC-V assembly for Computer Architecture education.

Assembly Educational

GitHub

numcEducational

NumPy-like library in C achieving 20x speedup using OpenMP, SIMD, and loop unrolling optimizations.

C 20x Faster

GitHub

Open Source

TruLens

Area Triager

Maintained by Snowflake · github.com/truera/trulens · Area Triager since September 11, 2026

Open-source LLM and agent evaluation and tracing library. 13 merged pull requests: one observability feature (streaming tokens, time to first token, and throughput in the OpenTelemetry path) and twelve correctness fixes across the database and persistence layer, dashboard, evaluation API, IR metrics, OpenAI and Anthropic providers, instrumentation, leaderboard aggregation, and Python 3.10 compatibility. Listed in MAINTAINERS.md with scope over OpenTelemetry instrumentation, evaluation and database correctness, and shared core code. Feature proposal #2741 (repeated-trial self-consistency for LLM-as-a-judge metrics) is grounded in the Coin Flip Judge paper.

MAINTAINERS.md Merged PRs

inspect_evals

Contributor

UK AI Security Institute · github.com/UKGovernmentBEIS/inspect_evals

The UK AI Security Institute's evaluation suite for Inspect AI. Merged PR #2317 migrating the shared metric helpers (mean, standard deviation, and standard error) to the framework's aggregation API; credited in release v0.20.0 (September 2026).

Release notes v0.20.0

Awards & Recognition

Peer Reviewer

191 peer reviews for 8 journals (ORCID, September 2026)

Engineering Applications of Artificial Intelligence, Image and Vision Computing, Neural Networks, Information Fusion, Neurocomputing, Expert Systems with Applications, Knowledge-Based Systems, and Data and Knowledge Engineering.

ORCID record

Elsevier Recognised Reviewer Certificates

2024 to 2026 · nine journals

Certificates awarded for review volume across the following journals:

Engineering Applications of Artificial Intelligence 109 Image and Vision Computing 33 Neural Networks 26 Information Fusion 18 Neurocomputing 3 Expert Systems with Applications 1 Knowledge-Based Systems 1 Data and Knowledge Engineering 1 International Journal of Intelligent Networks 1

GroundLM 2026 Workshop at EMNLP 2026

Budapest, 2026

"The Coin Flip Judge?" accepted as an Archival Long Paper (Program Chairs decision: Accept; two area-chair recommendations to accept).

UC Berkeley Skydeck Pad-13

2023

FibonAI accepted into UC Berkeley's Skydeck Pad-13 Incubator (competing against 5,000+ startups globally).