Aaron Kushner, Ph.D.

Bio-inspired materials scientist and inventor · now an AI-native builder of research systems, engineering programs, and public education.

akushnerphd.me · linkedin.com/in/aaron-m-kushner · aaron.m.kushner@gmail.com · Oakland, California

Statement

What made Jobs a stand-out success was not just intellect — it was also the combination of intuition and experiential wisdom he felt was an asymmetric innovation booster. I am Gen-X with a Ph.D. in Chemistry, and a documented record of exceptional non-linear/lateral thinking and intuition resulting in multi-disciplinary innovations — two materials patents and a self-healing polymer platform that came from asking what engineering materials problems nature already solved, and extracting the core reproducible nano-engineering design.

“Intuition is a very powerful thing, more powerful than intellect, in my opinion. That’s had a big impact on my work. Western rational thought is not an innate human characteristic; it is learned and is the great achievement of Western civilization. … the power of intuition and experiential wisdom.” (emphasis added)Steve Jobs, to Walter Isaacson · Steve Jobs, 2011 · p. 73–74

My last two years are effectively the experiment: one non-engineer, an AI workbench I built myself, and a set of programs that would each have required a funded team.

I also possess the skills, awareness, empathy, and experience required to lead a team. The era of Jobs, Musk, and Gates required a harsh personality to convert vision into shipped product. It no longer does. AI, harnessed properly ("Personal AGI"), supports and encourages the drive and focus that used to demand a difficult leader. The other brutal requirement is gone too: manually separating A-players from B-players by force of personal judgment. With a provenance-documented build and performance record at every seat, people self-sort on evidence. What remains scarce is the thing the machine cannot supply — taste, cross-domain intuition, and knowing which problem is worth solving... there too quantity/quality of output in terms of company (and personal/professional) value generation is now easily assessed.

AI practice — three and a half years (Git records available on request)

Dec 2022

Opened an OpenAI account seventeen days after ChatGPT was released to the public. First paid use February 2023. A research instrument at this stage rather than a build tool — expert-role prompting across chemistry, materials, and market questions. Interesting and incrementally efficiency enhancing.[NIH proposal available on request]

Mar 2024

ChatGPT Plus. The GPT‑4 era — daily working use across research, writing, and code, cut and paste to VS Code.[Skills: Python · scripted data handling · long-form document generation]

Apr–Dec 2025

ChatGPT Pro. Built independently, browser to VS Code:NormaSky — an air-traffic anomaly detector: a masked-autoencoder transformer trained on millions of rows of ADS-B state vectors to learn normal traffic and flag deviation from it, with a live dashboard; cloud-GPU training, drafted for publication.SmartColumn — predicts flash-chromatography conditions from molecular structure and emits them as vendor-specific method parameters for automated purification systems, with a post-run intake that returns observed results as correction data.Also a multi-aircraft collision-avoidance simulation with logged experiments and a web viewer, an adaptive Mandarin tutor with speech-signal analysis and per-user skill tracking, a literature-harvesting pipeline over journal supporting information, and Unity XR work. Versioned by copying folders and stamping the date — the tooling could not carry state between sessions.[Skills: Python · PyTorch / PyTorch Lightning · scikit-learn · pandas · NumPy · Streamlit · FastAPI · Flask · Docker · cloud GPU (A100) · Trino / Parquet · Supabase · Next.js · Unity/C# · Git · Vercel]

Dec 2025

Switched to Claude.

Mar 2026 —

First Claude Code session, 11 March. Everything measured below starts here.[Skills: LoRA fine-tuning (FLUX.1-dev) · pre-registered model evaluation]

The measured record

11,320prompts across 604 working sessions since 11 March 2026

101,387model turns in June–July 2026 alone

42.2 billiontokens of model context processed in that same two-month window

173 millionoutput tokens generated — drafts, code, analysis, artifacts

2,459commits since 30 March 2026, across 58 project directories

1person — no engineering staff, no external development budget

The practice is not prompting. It is systems design: parallel collaborating specialist agents with both deep knowledge and defined scopes, memory that survives session death, verification gates that block unevidenced claims, and a documented behavioral layer refined continuously against real failures.

Proof in results

Each program below links to its full description. State is current as of July 2026.

General Research & Development WorkbenchWorking prototype
The flagship. A multi-agent orchestration harness built on Claude Code: persistent memory across sessions, continuity protocols that survive process death, parallel specialist lanes coordinating through durable artifacts, and verification gates that block a claim of “done” without observed evidence. Not a wrapper — a behavioral layer accumulated from months of daily use. The version I run in my own IDE is a working prototype and is what produced everything below; the hosted VM workbench is in development; the mass-market layer that would make it operable by a non-technical user is conceptual but inevitable — and its target now has a market-legible name: Personal AGI, models and behavior uniquely tailored to each individual user. I built this harness to adapt to one person, and it worked: it carries my context, standards, and working style across sessions — the provenance record above is the measured proof. The institutional application adds: Tokenomics — the AI-spend measurement problem the business press says has no accounting practices yet. Every seat’s output flows to a higher-altitude seat as a synthesized report of the type, scale, and impact of granular token spend; the synthesis repeats up the chain — seat to team to corporate — with at least one human signing off and making every final call. Each claim is measured, per-seat, and traceable to the commit; the generator is push-button, and its first real report renders unsigned until a human signs it. Tokenomics is the investor side; the human side is the removal of manager–builder friction.

The AirPark Digital TwinBuilt · demonstrated
An aircraft program run by one person: layout and mass balance, CFD over the real airframe, structural FEM, flight simulation, automated engine-out landings, crash survivability. The pass/fail oracle is physics, so the loop closes without a human in it. The demonstration product of record for the workbench, and the seed of a universal atoms-to-objects-to-systems engineering engine. Pre-AI equivalent: roughly $250,000, a team of engineers and lawyers, one year. Actual: roughly $5,000 plus a subscription, one non-engineer, three months. The $250,000 still needs to be spent on a team of engineers and lawyers as it becomes a real-world business. But the investment risk is reduced dramatically.

Responsible AI DiffusionAuthored · verified · published
Independent policy research on maximum AI diffusion at maximum safe speed, communicated to the public (prototype). Three papers, each subjected to multi-model adversarial review before release and published with quantitative AI-use transparency — the contribution of the machine is disclosed and measured, not implied.

Provenance in the AI Age · AuthorAwarePlatform live
Author-side certification and a shared hash registry: tamper-evident records of how work was actually made, so authorship survives contact with generative tools.

TeacherAwareProduct complete · ready to deploy
The demonstration product of the provenance work, built out to completion: writing integrity for grades 7–12 that records the writing process and issues a tamper-evident credential, instead of inferring integrity from the finished product. Free and LMS-friendly.

The Citizen Scientist ExchangeMVP shipped
Verifiable independent research: three frontier-model reviewers per endpoint, tamper-evident by construction, human-decided at the editorial desk. Review capacity scales with participation rather than with the free time of a scarce referee pool.

open-edTwo products live, free
Adaptive, case-based open education — the matched school for any mind, at zero marginal cost. Shipping as free products while the K-12 engine matures.

1 + 1 → Attention Is All You Need · the LADDER curriculumCurriculum shipped, free
41 lessons from 1 + 1 to Attention Is All You Need, with an interactive map of the whole route, a teaching figure at every station, and the history of the person who found each idea — built for people who did not decide to pursue STEM, and people who prefer spatial reasoning to symbolic notation.

Mandarin Lab · Tone LabLive app
Speak the sentence and your pitch contour is scored per syllable against native renderings, normalized to your voice — tone sandhi handled. An AI tutor coaches you in Mandarin at your level; your skill map persists across visits.

The Bioengineering MachineStartup seed
The workbench pointed at biology: full-suite in-silico R&D, de-risking programs that once demanded eight-figure wet-lab spend.

FlyIRL · the SkyParkPre-launch validated
Over 1,000 signups at under $2 per lead in under a month; the AirPark digital twin is its engineering arm.

Selected pre-AI record

NSF Convergence Accelerator proposal — the PRE-VIS Kit · ScaledNano, Inc., with Cornell University and University of Dayton
Principal author of a trans-disciplinary ed-tech/av-tech crossover research proposal: a Parallel Reverse Engineering Virtual STEM kit using bio-inspired design — students reverse-engineer a bird and a “birdplane” in parallel — targeting 12–15-year-olds, the developmental window that most shapes STEM career interest and mathematical self-efficacy, with an explicit focus on underrepresented and underserved populations, domestically and in India. Full proposal available on request.

CEO / CTO, GK Materials · 2012–2015
Commercialized a bio-inspired self-healing polymer platform.

Postdoctoral Researcher, Guan Lab, UC Irvine · 2010–2012
Developed new project directions; mentored undergraduate researchers.

Research Associate, Advanced Medicine / Theravance · 1999–2003
Medicinal chemistry.

Consulting · Drywired LLC (nanocoating IP, 2020–2021) · GPH Private Equity (materials and materials-informatics investments, 2016–present)

Education

Ph.D., Bio-inspired Nanomaterials
University of California, Irvine · 2010

B.S., Organic Chemistry
University of California, Berkeley · 1999

Electrical & Mechanical Engineering
Laney College · coursework, self-directed toward the aircraft program

The College Preparatory School
Oakland, California · 1994

Patents & publications

Patents: Bio-inspired method to obtain multifunctional dynamic nanocomposites · Synthesis of multiphase self-healing polymers from commodity monomers.

Selected publications: Forced unfolding of single-chain polymeric nanoparticles · Direct correlation of single-molecule behavior with bulk mechanical properties for a biomimetic polymer · Supramolecular block copolymers for self-healing thermoplastic elastomers · Multiphase design of autonomic self-healing materials · Modular design in natural and biomimetic soft materials · A biomimetic modular polymer with tough and adaptive properties · Biomimetic design of a reversibly unfolding cross-linker to enhance mechanical properties of 3D network polymers.

Technical

Multi-agent AI orchestration and harness design · adversarial multi-model evaluation · multifunctional materials design · molecules-to-materials multi-scale simulation, electronic structure to mechanical properties · materials genome library and data pipeline design · real-world experiment design and data engineering for a practical materials-genome model over a useful proprietary slice of chemical space · conversion of laboratory research into IP · multi-step organic and polymer synthesis to kilogram scale · molecular characterization (NMR, IR, MS, X-ray, GPC, DLS) · materials characterization (DSC, TGA, DMA, Instron) · nanostructure characterization (AFM, SEM, TEM, SAXS, GI-SAXS) · machine learning · English (native), Mandarin (intermediate).