Skip to content
Ben Kennedy AI Engineer & Researcher · Kennedy Applied Sciences Research Group
Applied AI · Research to product

Ben Kennedy

Building intelligent systems that are efficient, understandable, and useful — peer-reviewed research, commercialized into tools for regulated, privacy-sensitive environments.

Human in the loop, by design.
01 · Bodies of Work

Four programs, one thesis.

Every project below is a different test of the same idea: intelligence belongs close to the person using it — measurable, explainable, and private. Open a program to read the full case study.

Case Study · Research Program 01The SCAN Ecosystem

Problem
Complex decisions deserve support that reasons like a team of specialists — not a single opaque model — while keeping a human accountable for the outcome.
Constraints
Cognitive plausibility (the architecture is aligned to function), human-in-the-loop oversight as a hard requirement, and claims that can be measured rather than asserted.
Decisions

SCAN anchors the ecosystem as the core cognitive architecture. SCANUE extends it into a multi-agent adaptive learning framework with oversight built in. SCANAQ closes the loop with measurement: a 36-item covering cognitive and affective self-regulation across eight subscales.

System
Perception context intake Specialist A planning Specialist B valuation Specialist C context Conflict monitor reconcile · weigh Response recommendation human oversight loop

1 / 4Perception

A decision context enters the system once and is shared, so every downstream specialist reasons over the same grounded picture instead of its own private copy of the problem.

2 / 4Specialist processing

Parallel agents — each aligned to a distinct prefrontal function such as planning, valuation, or contextual memory — work the problem independently. Disagreement between them is a feature, not a failure.

3 / 4Conflict resolution

A dedicated monitor reconciles the specialists: it surfaces where they diverge, weighs their confidence, and composes a coherent position rather than averaging everything into mush.

4 / 4Response, with a human in the loop

The system returns a recommendation, never a fait accompli. The human decision-maker stays in the loop by construction — oversight is wired into the architecture, not bolted on.

Evidence
Three peer-reviewed publications through Springer and IGI Global, and a dedicated research home at SCANERAD.com. SCAN · Springer 2025 SCANUE · IGI Global 2025 SCANERAD.com →
Outcome
A published architecture family with its own measurement instrument — and the conceptual backbone for the applied, on-device systems below.

Case Study · Research Program 02STAC & Neuromorphic AI

Problem
Transformer-class AI is extraordinarily capable and extraordinarily hungry. If intelligence is going to live on-device — in air-gapped, power-constrained environments — the computation itself has to get cheaper.
Constraints
Biological plausibility, compatibility with modern architectures rather than a clean-slate replacement, and a path that eventually runs on real hardware.
Decisions
Integrate spiking neural networks with transformer computation instead of choosing between them: keep the representational power, replace dense always-on arithmetic with sparse, event-driven activity where it counts.
System
Conventional pathway every unit computes at every step Spiking pathway computation only when events occur

Dense pathways burn arithmetic on silence. Spiking pathways encode information in the timing of discrete events, so quiet inputs cost almost nothing — the property neuromorphic hardware is built to exploit.

Evidence
Published as a chapter with IGI Global, with the mechanics explored hands-on in the spiking-network explainer. STAC · IGI Global 2025 Inside a Spiking Neural Network →
Outcome
A published framework for efficient, biologically plausible computation that feeds directly into the on-device research agenda at Kennedy Applied Sciences.

Case Study · Applied AIStratoSort

Problem
Professionals in defense, healthcare, legal, and finance drown in documents — and the one class of tool that could help, cloud AI, is exactly the tool their data can never touch.
Constraints
Zero data egress. Not "encrypted in transit" — zero: a fully self-contained, air-gapped runtime on ordinary hardware, usable by people whose job is not machine learning.
Decisions
Bring the intelligence to the data. Ship semantic search, retrieval-augmented generation, knowledge-graph visualization, and intelligent file routing as a desktop application that runs every model locally.
System

The gesture of the product in one motion: unstructured files become an organized, searchable knowledge structure — and nothing ever leaves the machine.

Evidence
A shipping desktop product, built for regulated environments. stratosort.com →
Outcome
Commercial validation of the lab's central bet: on-device intelligence is not a compromise — in regulated domains it is the only product that can exist at all.

Case Study · EducationInteractive AI Explainers

Problem
The mental models behind modern AI are locked in papers. Most people who use these systems daily have never seen an embedding move or an attention head attend.
Constraints
Browser-only, no accounts, no heavyweight dependencies, keyboard-accessible — and understandable in about thirty seconds of manipulation, before a single paragraph is read.
Decisions
Build instruments, not articles. Each explainer is a standalone interactive machine you operate directly; the prose exists to annotate the machine, not the other way around.
Outcome
A growing series of standalone teaching instruments — the public-facing, plain-language edge of the research programs above.
02 · Research Arc

One line of work, sustained.

Not a list of publications — a progression. Each step below builds on the one before it, from cognitive architecture to measurement to efficiency to systems in the field.

  1. Foundation

    PhD, Artificial Intelligence

    Capitol Technology University — following an MS in Information Technology Management and a decade delivering production technology in regulated, PHI-handling environments across healthcare and defense.

  2. Architecture · Springer 2025

    SCAN — Synthetic Cognitive Augmentation Network

    The anchor publication: a brain-inspired cognitive architecture for decision support, presented at SEET 2025 and published in Springer CCIS Vol. 2725.

  3. Adaptation · IGI Global 2025

    SCANUE — multi-agent adaptive learning

    SCAN extended into an adaptive multi-agent framework with human-in-the-loop oversight built in, published as Beyond Intelligence (IGI Global).

  4. Measurement

    SCANAQ — the instrument

    A 36-item psychometric instrument measuring cognitive and affective self-regulation across eight subscales — because an architecture that claims to augment cognition should be measurable.

  5. Efficiency · IGI Global 2025

    STAC — neuromorphic computation

    Spiking neural networks integrated with transformer architectures for efficient, biologically plausible computation, published as Aligned Minds, Efficient Machines (IGI Global).

  6. Open Experiment

    ItaSoRL — readout, not reward

    A tabula-rasa artificial-life experiment: a digital organism raised in a world that is sometimes authentic, sometimes a subtly flawed surrogate — asking when a mind comes to represent something it was never asked to care about. Open source, pre-registered, with a public research log and code.

  7. Now

    Current systems

    On-device inference, cognitive architecture, and privacy-first systems for air-gapped regulated domains — carried out through Kennedy Applied Sciences Research Group and validated in the products below.

03 · Commercialized Tools

Intelligence validated in the field.

Select tools developed through research, commercialized to test the underlying science in real-world regulated environments.

Research Practice

Kennedy Applied Sciences Research Group

Self-funded applied AI research practice. Privacy-first, on-device intelligence for defense, healthcare, legal, and finance. The most valuable data in these domains can never leave the device, so the intelligence comes to the data.

Product

StratoSort

A desktop application that brings modern AI to your files without sending a single byte off your machine. Semantic search, RAG, knowledge-graph visualization, and intelligent file routing, all running locally in a fully self-contained, air-gapped runtime.

Product

Odta

On-device task and time manager with local ambient intelligence. Semantic search, smart task classification, and duplicate detection, powered entirely by embeddings running in the browser. Free and open source, works offline, and no task text ever leaves the device.

Product

Rephrame

A private, offline-first cognitive behavioral therapy journal. Thought records, guided reframing, behavioral activation, and worry postponement, with no account, no server, and no tracking. Everything stays on the device.

Product

RSSE

A fully automated, rules-based trading engine for TradingView covering futures, stocks, and crypto, built to take emotion out of the trade. Every setup is qualified through a multi-stage signal process, sized risk-first, and managed to its exit, with prop-firm protections and hands-free webhook execution. Across a roughly 6-year backtest, $2,000 grew to a simulated +7,580% return, about 4× buy-and-hold.

Results are simulated and not predictive. RSSE is a research tool, not financial advice.

04 · Open Research

Artifacts and reference implementations.

Selected research code, experiments, and open-source contributions on GitHub.

Contribution activity
Loading contributions from GitHub
05 · Background

Where this comes from.

AI Engineer and Researcher with a PhD in Artificial Intelligence (Capitol Technology University) and three peer-reviewed publications. A decade delivering production technology in regulated, PHI-handling environments spanning healthcare and defense.

Through Kennedy Applied Sciences Research Group, I explore on-device intelligence, cognitive architecture, neuromorphic computing, and behavioral technology. Select tools are commercialized from this research to validate the science in real-world high-stakes settings. This practice is self-funded, operated independently of full-time employment, with no competing outside engagements.

  • PhD, Artificial Intelligence, Capitol Technology University
  • 3 peer-reviewed publications (IGI Global, Springer)
  • MS Information Technology Management
  • On-device & privacy-first AI
  • Cognitive architecture
  • Neuromorphic computing
  • Psychometric AI assessment
  • Knowledge management
  • Quantitative systems
  • Behavioral technology
Prefrontal cortex alignment SCAN's specialist agents mirror functions associated with regions of the prefrontal cortex — planning, valuation, conflict monitoring — so the architecture's division of labor has a cognitive rationale, not just an engineering one.
Psychometric instrument A structured, validated questionnaire designed to measure psychological constructs. SCANAQ uses 36 items across eight subscales to quantify cognitive and affective self-regulation.
Neuromorphic hardware Chips engineered to compute the way spiking networks do — asynchronous, event-driven, and massively parallel — rather than executing dense matrix arithmetic on a clock.
↑↓ navigate ↵ select esc close