• LearnOS - Self-hosted, open-source AI university: agents that build personalized roadmaps, tutor in real time, and grade toward mastery. One OpenRouter key, any model.
• rerun-bench - Replay your own repo's past bug fixes on your coding agent, several times each, and see where it is reliable and where it is a coin flip.
• nerf-watch - See what loads into every coding agent session before you type, cut what you never use, and catch silent regressions from your own logs.
• snap-back - Undo what your coding agent did, including what built-in checkpoints miss: overwritten .env files, deleted untracked files, and shell side effects.
• agent-fence - One permission policy for every coding agent: replay your existing logs against it, then block force pushes, secret reads and rm -rf for real.
• eval-builder - Is your LLM judge a coin flip? Tests it for position bias, flip-flopping and padding bias with no labels, then turns your real logs into an eval suite.
• paper-repro - Point your agent at a paper. It reruns the claim and gives you a verdict with receipts, or tells you exactly why it couldn't.
• train-doctor - Finds the bottleneck in a slow training run, tries the fixes, and keeps only speedups it can prove with the loss unchanged.
• browser-port - Turn an open-source C/C++ SDL game into a link anyone can play, in one command. Play the demos
• web-revive - Bring a dead Flash or Java web game back from its old URL as a playable link. Play the demos
• open-agent-lab - Independent, reproducible evaluation of coding agents: same tasks, many runs, intervals shown.

More: notebook-to-pipeline · launch-day-kit · Synthora · TableSage

• REA: Reverse Engineer Anything - Android DEX static values through the JADX bridge, Mach-O segment and section name parsing, adb integer intent extras, and precise input errors for byte entropy windows, regex patterns, plist symlink loops, gRPC-web decoding and CDP targets.
• Headroom - Keep indented lines intact in OpenAI whitespace normalization so cached prompts stay byte-identical.
• llmfit - Keep the registry VRAM reading on Windows when PowerShell exits non-zero.
• dbx - SQL Server max string and rowversion mapping for cross-database transfer, correct sorting of WITH query results, and clause keyword ranking in SQL completion.
• oMLX - Show remote rank output when a cluster rank exits, and answer DiffusionGemma's vision tower check in the oQ sanitize proxy.
• CodexBar - Show the Cursor X Subscription weekly pool as the main meter. Credited in the 0.79.1 changelog.
• TruLens (Snowflake) - Area Triager, listed in MAINTAINERS.md. OpenTelemetry streaming metrics (TTFT, throughput), Jury reliability metadata, and correctness fixes across the database layer, OTEL instrumentation, providers and dashboard.
• inspect_evals (UK AI Security Institute) - Migrated metric helpers to inspect_ai's aggregate() and unscored reasons to Score.reason, plus eval scoring and docs fixes. Credited in releases v0.20.0, v0.22.0 and v0.25.0.

class AbelYagubyan:
    def __init__(self):
        self.current_role = "Ph.D. Student in Artificial Intelligence @ University of Georgia"
        self.building = "LearnOS, an open-source agentic AI university"
        self.location = "San Francisco, CA"
        self.education = {
            "phd": "Ph.D. Artificial Intelligence @ University of Georgia (2026 - present)",
            "masters": "M.S. Computer Science @ Northwestern (Summa Cum Laude)",
            "bachelors": "B.A. Computer Science & Applied Math @ UC Berkeley",
        }
        self.research = [
            "LLM-as-a-judge reliability",
            "LLM agent reproducibility",
            "Evaluation science",
            "Open-source LLM evaluation tooling",
        ]

    def achievements(self):
        return {
            "accelerator": "Y Combinator, Spring 2026 batch",
            "papers": "4 published (GroundLM @ EMNLP 2026, MNRAS 2022, ACM SIGCSE 2022)",
            "citations": "45, incl. Meta Superintelligence Labs and Alibaba",
            "peer_review": "196 reviews across 9 Elsevier AI journals",
            "open_source": "Area Triager @ TruLens (Snowflake), contributor @ UK AISI inspect_evals",
            "industry": "Senior Data Scientist @ C3.ai, Co-founder @ FibonAI (Berkeley SkyDeck)",
        }

Peer review: 196 verified reviews across nine Elsevier journals, including Engineering Applications of Artificial Intelligence, Image and Vision Computing, Neural Networks, and Information Fusion. See my ORCID record.

Spring 2026 Batch Startup accelerator	Archival Long Paper Sole author, Budapest	Elsevier Recognised Reviewer Nine AI and vision journals
Area Triager Snowflake, 30 merged PRs	M.S. Computer Science Northwestern University	Research Contributor UPC++ performance, Pagoda Project

gantt
    title Professional Journey
    dateFormat YYYY-MM
    axisFormat %Y
    section Research & Education
    Ph.D. AI, University of Georgia   :active, 2026-08, 2031-05
    M.S. CS, Northwestern             :2022-09, 2023-06
    LBNL Research Contributor         :2022-06, 2023-04
    B.A. CS & Applied Math, Berkeley  :2018-08, 2022-05
    section Industry
    Y Combinator (Spring 2026)        :2026-03, 2026-06
    Senior Data Scientist, C3.ai      :2024-05, 2026-03
    FibonAI Co-founder                :2023-06, 2024-02
    Apple x UC Berkeley Intern        :2021-05, 2021-08

I'm always interested in collaborating on:

• ⚖️ Evaluation science: making LLM judges and agents measurably reliable
• 🧩 Open-source LLM evaluation and observability tooling
• 🎓 Agentic systems for education
• 💡 Early-stage AI startups