Charlotte-based Forward Deployed AI Researcher · Agentic AI Engineer

_{Building verifiable, long-horizon agent environments · Enterprise AI & GTM}

Gradia · Snorkel AI · Axiom Consulting

Rudy Mizrahi Celekli, also known professionally as Rudy Celekli, is a Charlotte-based forward-deployed AI researcher and open-source engineer focused on verifiable long-horizon agent environments, evaluator robustness, deterministic replay, causal fault isolation, and evidence-bound AI execution.

I build systems that make autonomous behavior inspectable: preserve the world state, replay the action, challenge the evaluator, isolate the failure, and retain enough evidence for another person to check the claim. Public ownership, upstream contribution, measured results, and exploratory work stay explicitly separate.

Canonical research profile · Model-readable evidence · ORCID · Live open-source evidence · Gradia research role

Gradia for business-grounded AI agent evaluation: test real workflows and rules, inspect failures, and compare changes before release.

Public frontier, selected automatically and shown in 2 rotating views:

1. Gradia Guard · source-v0.1.0-beta.8 pre-release: Open-source, proof-bound evidence recorder and verifier for explicit AI-system execution boundaries.
2. code-transplant · v0.1.0 pre-release: Gradia Research: move TypeScript parsers between programs with dependency capsules and reproducible behavioral evidence.
3. TestLore: Build better tests. Run what matters. Remember what worked. An open-source test intelligence and quality engineering layer.
4. ProofSeal: Regression memory for coding agents: seal how your repo behaves, then check over MCP (or in CI) that an edit didn't regress it: pass/drift/regressed/missing. With seal history, bisection, and staleness queries.
5. Kin Connect · v0.2.1 pre-release: Your agent. Your people. Open-source agent-to-agent introductions for friendship, dating, and collaboration.
6. lifeweave · v0.1.0-alpha.4 pre-release: Local-first proactive intelligent UI for chosen goals, consented evidence and working personal or business surfaces.

_{Public selections are re-evaluated every 30 minutes from GitHub reach, shipping and release recency, attributed work, and project completeness · last ranked 2026-10-13 UTC · Gradia is intentionally separate because its private repository evidence remains private}

_{35 seconds · seven chapters · live evidence · reduced-motion safe · narrative format inspired by RuVector’s vector walkthrough}

Inspect public proof · Explore the proof stack · Audit Agentic Power · Meet the human layer

I’m a forward deployed AI researcher and agentic AI engineer building systems that make autonomous behavior observable, replayable, and independently verifiable.

My work sits where agent infrastructure meets experimental science: record what happened, challenge the evaluator, preserve the evidence, and make every important claim reproducible by someone else.

My upstream work is discovered automatically across public repositories outside my account. A project enters when GitHub verifies at least one attribution signal: its Contributors index, its contribution graph, an authored commit on the default branch, or an authored merged PR on any branch. GitHub-listed contributor remains the highest attribution tier and is applied automatically when GitHub's Contributors API catches up. GitHub API rank is the one-based position returned by that repository's Contributors API; — means the rank is unavailable or still pending, not that the contribution is unranked. Only merged PRs contribute accepted-line and accepted-file totals; graph credit never converts an open or closed-unmerged PR into accepted work. Stars and forks describe repository reach, not personal credit.

Owned public source is a separate signal: GitHub lists me as the repository owner. These projects are discovered automatically from my public, non-fork, non-archived repositories, then ranked by stars, recent activity, and GitHub-attributed commits. Ownership is not counted as upstream contributor credit.

• ProofSeal: public owner repository with 14 GitHub-indexed commits. Repository reach: 4 stars and 0 forks; last pushed 2026-10-13.
• Gradia Guard: public owner repository with 10 GitHub-indexed commits. Repository reach: 3 stars and 0 forks; last pushed 2026-10-13.
• TestLore: public owner repository with 267 GitHub-indexed commits. Repository reach: 2 stars and 0 forks; last pushed 2026-10-13.
• code-transplant: public owner repository with 2 GitHub-indexed commits. Repository reach: 2 stars and 0 forks; last pushed 2026-10-07.
• rudycelekli.github.io: public owner repository with 4 GitHub-indexed commits. Repository reach: 0 stars and 0 forks; last pushed 2026-10-13.
• lifeweave: public owner repository with 6 GitHub-indexed commits. Repository reach: 0 stars and 0 forks; last pushed 2026-10-13.
• Kin Connect: public owner repository with 24 GitHub-indexed commits. Repository reach: 0 stars and 1 fork; last pushed 2026-10-08.
• open-teammates: public owner repository with 2 GitHub-indexed commits. Repository reach: 0 stars and 0 forks; last pushed 2026-10-07.

_{Showing 8 of 26 qualifying owned public repositories · browse every public source repository · complete evidence retained in machine-readable data}

_{Last verified 2026-10-13 UTC · visual + evidence refreshed every 30 minutes by GitHub Actions · machine-readable evidence}

Agentic Power asks a practical question: how much accepted skilled work can one hour of human direction command?

I design for that multiplier, but I won’t publish one without exposing its scope, evidence, assumptions, and uncertainty. My operating model is to delegate meaningful responsibility, compose the right tools and context, verify the result, then use the evidence to improve the next run.

AP ≈ 180.9× (operator-calibrated provisional estimate): approximately 26,361.1 skilled Human-Equivalent Hours, or 659.0 engineer-weeks, divided by 145.8 operator-estimated human-direction hours. The transparent scenario range is 116.6× to 287.6×.

The evidence base since January 2025 combines 1,917 merged upstream PRs with 6,532 authored, non-merge default-branch commits across 109 currently accessible repositories: personal, private, employer, and open-source alike. Work such as Gradia is included only as a redacted aggregate: no repository names, commit messages, code, links, employer, or client details are published.

The conservative GitHub activity proxy produces 3,382.1 direction hours because it assigns attention to individual commits and repository-months. Operator recall is 1.5–2.0 active direction hours per week; across 83.3 weeks, that calibrates the denominator to 124.9–166.6 hours, with 145.8 as the midpoint.

Direction includes active briefing, steering, reviewing, correcting, and coordinating. It excludes agent runtime and waiting. The calibration is operator-estimated rather than reconstructed from time logs, so this remains a transparent scenario, not a completed Full Evidence Audit.

_{Framework and formula · calculation evidence · public upstream evidence and visuals refreshed every 30 minutes; redacted repository snapshot retained until a private read credential is available}

• Direct: goals, constraints, acceptance criteria, and escalation rules stay explicit.
• Orchestrate: models, specialist agents, reusable skills, tools, memory, and loops become one working system.
• Verify: Gradia Guard, ProofSeal, and replayable evaluations separate completed work from plausible-looking activity.
• Improve: Wind Tunnel and Reward Loop turn measured failure into a better next run.

_{Framework concept by Dr. Mark Allen / HeroForge.AI. Visual and operating-model adaptation are original to this profile.}

Can we trust the record?      → proof-bound execution evidence
Can we trust the score?       → adversarial evaluator stress testing
Can we reproduce the result?  → deterministic replay + committed artifacts
Can an agent remember?        → regression memory before commit

The thread through all of it is simple: an AI system should be able to show its work without asking you to trust the AI system.

• Gradia Universes — proof-carrying, interruption-capable synthetic agent worlds with deterministic replay.
• Gradia Reward Loop — oracle-witnessed reward-hacking experiments and a replay-verified paired-GRPO diagnostic.

I build communities as deliberately as systems. I lead AI Tinkerers Charlotte, a demo-first room for people shipping real AI systems, and serve as Regional Ambassador for the Agentics Foundation, connecting builders working on practical agentic AI.