I design and harden the benchmark tasks used to evaluate how well AI agents actually reason, not just how well they pattern-match. My work sits at the intersection of software engineering, applied mathematics, and adversarial test design — built on a mechanical engineering foundation that keeps me grounded in first-principles problem solving.

profile = {
    "name"        : "Deepak Kumar",
    "aka"         : "Saranay",
    "location"    : "New Delhi, India",
    "focus"       : "AI benchmark design & evaluation engineering",
    "current_work": [
        "Authoring & hardening Harbor-format benchmark tasks — Project Dynamo (Handshake AI)",
        "PhD-level evaluation task design — Vedron.ai (astrophysics, gravitational-wave physics)",
    ],
    "also_building": [
        "Engineerspedia — engineering explainer blog",
        "Hindi-language physics & science YouTube channel",
    ],
    "education"   : ["B.Tech, Mechanical Engineering — RGPV", "PGDCA — SGSU"],
}

Benchmark & evaluation engineering — Project Dynamo (Handshake AI) I build and rework Harbor-format tasks that stress-test AI agents across scientific computing, machine learning, mathematics/formal reasoning, and physics simulation. Each task is engineered end-to-end: reference solution, isolated verifier, seeded randomness, and protected ground truth — designed so difficulty comes from genuine algorithmic and mathematical complexity, not ambiguity or information hiding. I work the full fork → PR → automated multi-gate review (tier1, QC gate, adversarial review, pass@k) → iterate cycle, and take on rework assignments to fix flagged tasks against reviewer findings.

AI evaluation — Vedron.ai Author PhD-level evaluation problems in astrophysics and gravitational-wave physics (magnetar energy budgets, strain/braking-index derivations) used to assess frontier model reasoning.

Science communication Run a Hindi-language physics/science YouTube channel producing short-form and long-form explainer content — including full production pipelines (narration scripting, Manim-animated storyboards, subtitled renders) — and write for Engineerspedia, an engineering-explainer blog aimed at making core engineering concepts accessible without the jargon.

• Adversarial task design — reference-solution engineering, isolated/protected verifiers, seeded determinism, anti-cheat isolation
• Formal & mathematical methods — Dafny formal verification, CP-SAT / OR-tools constraint solvers, PSLQ-based symbolic computation
• Infrastructure & CI/CD — Docker-validated pipelines, GitLab CI/CD with automated deploy + rollback, Kubernetes deployment
• Applied engineering — mechanical engineering background (RGPV) informing systems and simulation-based task design

• B.Tech, Mechanical Engineering — RGPV
• PGDCA — SGSU
• Prior roles: Subject Matter Expert at Chegg India, plus technical/content work at Trivium Services, Contentwhale, and Outlier
• Career path: mechanical engineering → technical content & subject-matter expertise → AI model training and evaluation