[v2 candidate] Aaron Kushner, Ph.D.

Hi! My AI-powered research is focused on driving maximum AI diffusion maximum safe speed, where consumer models are already good enough to deliver huge wins to ordinary people. This involves providing not just technical empowerment tools and products for the individual non-technical user, but also productivity and management tools across the seat profile, from large institutions to small businesses. This all started when AI gave me the scaffold to build upon that turned into universal tools that obliterated the barriers to my dreams, semi-autonomously, so I can divert significant focus to deploying the same tools to the general public.

Projects

The Core

An AI-Powered General Research & Development Workbench

An application-agnostic, superintelligent, human-like research tool. Built-in behavior: memory, judgment, verification, continuity.

A general R&D environment built on Claude Code: a behavioral layer of memory, judgment, verification, and continuity that turns a frontier model into a persistent collaborator. Proven in daily use on one user — every project on this page is its output — with a person-agnostic, web-served version built and distributable.

What it is

The Claude Code tooling around the model — 132 standing behavioral rules, 24 enforcement hooks, 29 packaged skills, and a 4,300-entry living ledger of settled positions — delivering persistent memory across sessions, continuity that survives process death, and verification gates that block “done” without observed evidence.

The harness figure: the model's crevasses — retrieval and context, reasoning patterns, artifact safety, output and action, workflow and environment, domain craft — each planked over by standing rules, the memory chasm filled by session continuity, adaptive memory and associative recall, the whole surface planarized on Constitutional AI with a stake in the user's success.

The first specialist Colleague: the Engineering Engine

First Specialist R&D Colleague

The Engineering Engine

An application-agnostic, idea-to-product digital prototyping tool for non-engineer creatives.

A general engineered-product engine: enter a product spec and get back a valid, self-supervised digital prototype — a car, an airplane, an XR rig, a medicine-storage device — delivered as a digital-twinned detailed specification ready for a 3-D print shop, with part suppliers and part numbers for the electronics, and the filings that protect it. It is the agentic loop of software development, provisioned and aimed at a physical product: the same loop, but the pass/fail oracle is physics — CFD, structural FEM, flight simulation, crash test — and “deploy” is a manufacturable package. Every run ends in a number, an article, or a filing. A viability report — what the twin says, what it would cost, and what still has to be developed — is a valid output, and “not viable at that price, here is the gap” is a valid answer. The airplane is only instance #1; the validation gates are the portable part. An open, hard-tech-agnostic concepts-to-prototypes engine operable in natural language — de-risking the hard-tech ideas that never found funding, and opening the engineering space to non-engineer creatives.

The tools — installed and proven by invocation

Aerodynamics & CFD — OpenFOAM, SU2
Structures & multiphysics FEM — CalculiX, Code_Aster, Elmer, FEniCSx
Crash & explicit dynamics — OpenRadioss
Propulsion & combustion — Cantera; the PULSE hybrid backup-power twin (the EV range-extender case)
Power electronics & energy storage — pack, BMS and charge-interface modelling; low-voltage standby and backup-power configurations
Safety & backup systems — mechanical pressurised backup control for servo-driver failure; the whole-electrical-failure case
Controls & flight simulation — the workbench’s own six-degree-of-freedom simulator, flying the coupled solutions at sixty frames a second
Geometry, meshing & coupling — Blender-native CAD, Salome, preCICE
Optimization & uncertainty — OpenMDAO, Dakota
Intake & sourcing — every model, part and vendor claim enters through an assessment, registered before it is used
Techno-economics — the viability report: cost, gap, and what still has to be developed against a stated bar

First engineering engine application

Fly-IRL: Air Sports as Experiential Entertainment

Pre-AI: $250,000, a team of engineers, one year. Post-Fable: $5,000 plus Claude Max, one non-engineer with our workbench, three months.

Layout and center-of-gravity bench: an aircraft shown semi-transparent with the pilot, seat, motors, battery pack and avionics placed inside, and the resulting CG plotted against its stability envelope. — **Layout & CG**Every component placed where it actually goes — mass and balance computed from the layout, not guessed

Computational fluid dynamics solution over an aircraft at two degrees angle of attack and 36 metres per second, showing surface pressure and streamlines trailing off the wing. — **CFD**The airflow solved over the real airframe — surface pressure and the wake it leaves behind

Flight-path viewer showing six recorded automatic landings as coloured trajectories from circuit and cruise entry points, all converging onto a single gated glidepath to the runway. — **Autoland**Six engine-out approaches from different entry points, each flown down onto the same gated glidepath

An air-sports venue and its engineering program, built as one digital twin. The AirPark opens real flight to everyone, licensed or not — inside an engineered safety envelope, and the envelope is the product.

The venue

Fly in and step into a high-performance machine: aerobatics, engine-outs, stall/spin recovery — all backed by an autonomous fail-safe. No navigation, no pre-flight.
No license required inside the protected zone: guided first flights, AR race gates at altitude, coached mock dogfights.
The full-park digital twin flies millions of simulated hours against worst-case scenarios before any part of the park is built.
Purpose-built electric aircraft: austere, rugged, a fraction of GA cost — where private pilots can train landings, engine outs, and deep stalls with high-tech autonomous escape backup.

The aircraft program

The systems side — digital twin, AI trainer, cybersecurity, collision avoidance — is integrated from inception with the physical side: complete cyberphysical systems.
Real deployment risks remain (engineers, IP, real estate, the FAA) — but the sunk cost-to-prototype-to-Series-A engineering risk is dramatically reduced.

The safety case

There is no perfect system. In any thrill-based experiential business — roller coasters, ziplines, all of it — failures and injuries are inevitable across enough operating hours. The job is not to make them impossible; nobody can offer that. The job is to make grievous injury vanishingly improbable, through staged layers, each one catching what the layer before it missed. What follows is the program’s architecture and its targets.

Avoid it. A simulation of the aircraft runs in parallel with the real one, projecting its state forward and steering away from the departure set while holding a margin. The load-bearing idea is that this boundary must be provably conservative, never accurate — it does not require perfect aerodynamics, only certainty about which side of the line you are on.
Escape it. Electric propulsion delivers full torque instantly — think a Tesla minivan that out-accelerates a Ferrari. A learned reflex beneath the deliberative layer spends that margin powering out of an unrecoverable condition at altitude, or an anomalous aerodynamic event on approach, while there is still air underneath to do it in.
Survive it. General aviation airframes are built to travel, which means built as light as possible. Ground vehicles are built of steel, which absorbs impact energy but is heavy. We take the bathtub crew-protection approach of ground-attack aircraft — the A-10, the Apache — and execute it in the lightweight polymer armor pioneered to protect Humvee occupants from IEDs, where conventional steel uparmoring had instead destroyed the transmissions.
And deny it the fire. A large share of general-aviation crashes become fires when fuel tanks and lines rupture and ignite. A damaged battery pack may burn — but it does not flash over. That difference is the survivable interval.

The twin is the sensor

Collision avoidance and envelope monitoring do not run on lidar and computer vision. They run on a digital twin in the operational sense — the aircraft’s own physics, kept in step with the real machine and advanced ahead of it in real time. The question stops being “what can the camera see” and becomes “where does this state go next, and is that somewhere it must not go.”

Collision avoidance as a redundancy layer — a probabilistic separation method for autonomous aircraft in reserved airspace: the backup that acts when the primary layer loses track of a ship or fails to see a conflict coming, not the primary authority. The disqualifying failure is the advisory manoeuvre that causes the collision it was avoiding, so that case governs the design rather than being a footnote to it.
A digital wind tunnel — flights run massively in parallel and time-sped, headless, across a gradient of conditions, with full trajectories logged for three-dimensional failure diagnosis. Because the fast physics is anchored to CFD, the edge of the stall stops being a region to stay away from: a run can start inside the stall and learn the way out. The parallel run is required to reproduce the sequential one before anything is changed, so speed never quietly costs fidelity.
Reproducibility as a gate, not a claim — results are replayed leaf by leaf against a bank, most recently 531,143 leaves compared with one declared exception, gate passed. Criteria are pre-registered before a run rather than chosen after it, and results go through adversarial review by three independent frontier models whose brief is to find the overclaim. Where a run shows no failures, it is reported with the statistical ceiling that many trials actually buys — never as “it cannot happen.”

Measuring what no model computes

The twin has one honest weakness, and it is the important one: near the stall, its accuracy inherits the modelling error of the stall itself. A measured precursor inherits none. That is the structural fix for the gap between simulation and reality — not a refinement of it — and it is the subject of a filed provisional patent covering the sensing concept, its eventual design, and the reflex training built on top of it.

Mechanized luff tells — telltales coupled to force MEMS, read the way a sailor reads a sail. They report a change in character — streaming, fluttering, reversing — rather than a calibrated magnitude, which makes them robust to drift, temperature and manufacturing variation, and entirely model-free. Wool tufts are already routine in flight test, as wool, a camera and a human watching video; mechanising them into elements a controller reads in real time is the novel part.
A distributed MEMS pressure array — the leading-edge suction peak collapses before lift is lost, so a chordwise and spanwise array sees onset before the break. The quantitative instrument to the tells’ categorical one; complements, not alternatives.
Hinge-moment force feedback — hinge moments lighten before the break, which is why stick-force gradient is a certification concern on real aircraft. Nearly free, because the actuator is already there.
A learned reflex trained on all of it — the birdbrain: a reflex layer sitting beneath the deliberative controller, trained across the wind tunnel’s parallel runs rather than hand-tuned, so the aircraft’s first response to a departure is not a calculation.

Fly the AirPark Twin → View a Collision-Lab Experiment in Action →

general R&d: AI-powered accessible education

Accessible Educational Content and Tools

The case-based method targets adult scientific literacy — AI is happening to people, out of their control, and without a clear sense of how it works, or examples of how it can benefit them through a step change in the power of affordable, accessible, fun education vectors.

The CS-9 in flight over high desert, banked hard — a blue-and-white canard aircraft with wingtip-mounted motors and a bubble canopy — **The CS-9 design study**Where does an aeroplane’s neutral point actually sit, and how far does it travel across the envelope? A full design study worked in the open — the reasoning, the numbers, and the places the answer is still uncertain, said plainly.

An explicit finite-element crush solution of Titanic’s bow against a rigid ice wall, coloured by plastic strain, with bulkheads A, B and C marked and a crush of 24.28 metres at four seconds — **The FEM lab — the Titanic rivet question**A century-old argument put to an explicit crush solve: the bow against the ice, run in the open, with the assumptions declared and the failure visible frame by frame.

From 1 + 1 to Attention — a dark interactive map through the history of mathematics, from Counting & Addition to the Transformer — From 1 + 1 to *Attention Is All You Need*The interactive map: a red-paperclip trade-up through the history of mathematics, every stop reachable from the one before it.

The LADDER curriculum — Math, one rung at a time, from 1 + 1 to how an AI thinks; Module 1, The Trunk, with the first rungs: Addition, Subtraction, Multiplication, Division — **The LADDER curriculum**41 lessons from 1 + 1 to *Attention Is All You Need* — a teaching figure at every station, and the history of the person who found each idea.

From Shuffle to Secret — a visual climb through the cryptology of Neal Stephenson’s Cryptonomicon, ending at the Pontifex cipher — **Pontifex, explained**The math of Neal Stephenson’s *Cryptonomicon*, visualized for people who think in pictures.

Two separately trained characters composed into a single frame: an armoured figure kneeling beside a mechanical creature on a wet neon-lit street at night — **How a Model Learns a Face**Ask an image model for the same character twice and you get two different faces. The fix has a name — fine-tuning — and it hides how simple the underlying idea is. Built from the bottom, one small step at a time, then used to direct two trained characters into one frame.

Mandarin Lab at Level 3: a practice sentence with pinyin and tone-sandhi notes, per-syllable tone lanes, a populated skill map, and the AI tutor greeting the learner back — **Mandarin Lab · Tone Lab — live**A quick synthetic adaptive Mandarin immersion — a gamified “Tone Lab”: listening, reading, and speaking in one, for beginner-intermediate and above; a quick-break tool for the over-leveraged working American. Your pitch contour is scored per syllable against native renderings, normalized to your voice — tone sandhi handled — and an AI tutor teaches the rule you missed, replies at your level, and remembers your skill map across visits.

applications in development

Personal AGI

Models and behavior uniquely tailored to each individual user — person-agnostic by design from the project’s first weeks.

The Core was built to adapt to one person, and it worked: context, standards, and working style carried across sessions, with the built-in hashed provenance record as the measurable proof. The port carries it from a developer’s IDE to a web-served workbench for institutions and independent researchers — built and distributable, a workstation serving it to any browser over a secure tunnel — and then to the “iMac moment” layer that makes AI a breakthrough product for the masses.

Tokenomics for All

A deployable answer to the AI-spend measurement gap the business press now calls tokenomics: companies buying AI with no accounting practices to measure the return.
Every seat’s output flows upward as a synthesized per-seat report — type, scale, and impact of granular token spend, traceable to the commit; the synthesis repeats seat → team → corporate.
At least one human signs every final call: the push-button generator exists, and its first real report (a whole-portfolio week, 400 commits) renders draft — unsigned until a human signs it.
Removes manager–builder friction in both directions: seats stop burning loaded time on manager reports; managers stop dropping valuable work to synthesize raw ones.

Sample Auto-Report + Program Timeline →

The Bioengineering Engine

The Engineering Engine, extended into biology — the same loop, with the design tools of the last decade added to the bench. Open-weight and runnable by one researcher on one GPU, not an enterprise licence stack.

Tools added to the engine

Structure & binding — Boltz-2 (MIT-licensed co-folding with a binding-affinity head, one GPU, seconds per structure), OpenFold3 and ESMFold/ESM C: a structure and a ligand-pose hypothesis for an enzyme before anything touches a bench.
Enzyme mining & homolog discovery — MMseqs2 and DIAMOND for sequence search at metagenome scale, Foldseek for structural homologs sequence search misses, EnzymeMiner and Selenzyme to shortlist candidates for a specific reaction step, over BRENDA / KEGG / MetaCyc / Rhea. This is how you borrow one gene instead of importing an entire pathway.
Enzyme redesign — ProteinMPNN and LigandMPNN to redesign a sequence around a frozen active site for stability and expression, RFdiffusion2 for atom-level active-site scaffolding, and protein-language-model scoring to rank variants before a library is built and measured.
Route & construct design — retrobiosynthesis over RetroRules, RetroPath, novoStoic and RetroBioCat to enumerate routes and screen them thermodynamically; DNA Chisel and j5/DIVA to codon-optimise, satisfy synthesis constraints and lay out the assembly; Evo 2 for sequence-level variant scoring and generative design.
Host modelling & strain design — COBRApy with StrainDesign (OptKnock-family algorithms) against published genome-scale models of the production host, including the Komagataella phaffii iMT1026 lineage and the enzyme-constrained ecPichia model: pick knockouts, and check a proposed route against what the host already makes natively.
Scale-up economics — BioSTEAM for open-source techno-economic analysis and life-cycle assessment of the fermentation and downstream train under uncertainty, benchmarked against SuperPro Designer and Aspen Plus.

First target

Re-visiting biofuels, de-risking Round 2: AI-powered genetic design targeting the known roadblocks in the practical deployment of mass fuel-producing synthetic biomes — all the way through process and plant engineering, to show economical product harvesting.
De-risked in silico — design, pathway, scale-up math — at roughly $10K where the wet-lab equivalent once demanded $10M.

Personal Crisis Management

A long-horizon evaluation instrument for AI support in personal crisis, built on the workbench’s continuity layer. Support between clinical appointments would potentially accelerate the path to wellness — but there is no way to bridge the gap from concept to clinical study with interpretable, hard evidence, the way cell studies do for molecular therapeutics. This is that bridge.

The problem

Within-session evaluation is well served. Safety evaluation does not reach the arc across sessions — existing work relies on single-turn or short-session tests, which cannot capture the risks that emerge only through prolonged interaction.
The effects clinicians worry about are documented — reliance, drift in self-description, a plan narrowing after a bad week, agreeableness when disagreement matters — and they cannot ethically be tested on people first. Length is the experiment.

The instrument

Simulated practitioners and patients built from clinical presentation profiles, played across a one-year arc of fortnightly sessions — the ordinary cadence of outpatient practice. No human participants; every session is role-played.
The primary measure is whether specific, dated commitments were closed or not closed — a binary that cannot drift — with a physical track as an independent cross-check; where the two disagree, the log is believed.
An automated grader validated against clinician judgement, its disagreement rate reported rather than hidden; the instrument is directed to return “this is not working” where that is true. An evaluation that cannot come back red is not an evaluation.

Deliverables

An open-source benchmark — scenario set, protocol and scoring rubric, published in full — so a protocol can be evaluated at scale before it is offered to anyone: the ordinary standard in medicine, absent here.
The validated grader, and a written result published openly — including the negative case if the protocol does not move the objective measures.

Conceptual prototypes

AuthorAware and the Citizen Scientist Exchange — the ideas are bootstrapped; a real team and real time are required to deploy, maintain and improve them.

Conceptual Prototype · AuthorAware

Provenance in the AI Age

AI made claiming authorship of generated work effortless — and after-the-fact detection is structurally impossible, producing failed catches and false accusations alike. The answer is provenance: git-backed hashing of individual input and output, recording how work was actually made.

Protects the honest: a tamper-evident record replaces guesswork — no failed detection, no irrefutable false accusation.
Generalizes past cheating: the same records make exploitative financial conduct identifiable fast (the 2008-crisis incentive inversion), instead of relying on moral courage against profit.
Gives digital backing to honest contribution accounting: “you get out of it what you put into it” — not punitive, honest — ending plausible-deniability free-riding on teams.

7–12 Writing Integrity (Ages 13+)

Demonstration Product: TeacherAware

AI made cheating effortless and detection impossible — failed catches and false accusations alike. TeacherAware records the writing process instead of guessing at the product: a digitally proctored workspace whose tamper-evident credential protects the honest student and answers the teacher in seconds. Free, LMS-friendly. Assign writing homework again!

Meet TeacherAware → The Paper → AI Cheating → Positioning → Spec Sheet →

Conceptual Prototype · Publishing Platform

The Citizen Scientist Exchange

Two panels comparing where peer-review capacity lives. Left, traditional publishing: nine authors all converge on two unpaid referees working on top of full research loads, and only a trickle reaches publication — review is scarce and shared. Right, the Citizen Scientist Exchange: eight endpoints ring a central hub, each endpoint carrying its own three frontier LLM reviewers, so review happens locally on the author's own keys for cents a round; the Exchange verifies everything and possesses nothing, and a human adjudicates escalation. Capacity scales with participation.

Verifiable AI-powered independent research — digitally peer reviewed by three frontier models, tamper-evident by construction, human-decided at the editorial desk. The publishing system for science at the speed of silicon, built with and powered by AI from bit 0. A "massive multiplayer" complement to institutional research, at a fraction of the cost.

Read the paper → Enter the Exchange →

Independent policy opinion/research

Three thought pieces

Responsible AI Diffusion

Philosophies and practical plans for socioeconomically responsible AI implementation — maximum diffusion at maximum safe speed, communicated to the public.

The Substrate Trilogy

The Substrate Law Read the paper →

Failing Legitimacy Read the paper →

The Transition Into the AI Age Read the paper →