Reuven Cohen – Cognitum (Live Agentic Engineering) (Talk Outline)

A Craft 2026 end-of-day live session by Reuven Cohen (Toronto; ~30 years building software, started very young; ~4 years “vibe coding” since mid-2022). Rather than a two-hour presentation, he does live vibe coding / agentic engineering and walks through his stack, then hands off to his partner Nick for demos before the Q&A. Format: screen-shared live demo (built in Lovable) with a running narrative. The talk’s own structure: the thesis (“if agents are the workflow, silicon is the runtime”) → the Roo Stack harness (Roo Flow, formerly Claude Flow): plugins, ADRs, CI guard, meta-packages, London-school TDD → the origin story (expensive swarms → Anthropic’s flat plan → Claude Flow’s overnight success → consulting business → the Agentics Foundation) → Cognitum: contrastive/null-space physical-world intelligence, its use cases, the Rust/WASM/NAPI-RS engineering, min-cut breakthrough, and investors → Nick’s demos → Q&A.

Note: the “Nick” here is a Cognitum / Agentics Foundation partner, not the Nick Brown of the capacity-planning talk.

1. Thesis — “If Agents Are the Workflow, Silicon Is the Runtime”

1.1 The idea he’s been working through for a year

• ”If agents are the workflow, then the silicon is the runtime” — a founding thought for the company.

1.2 Anyone can now build anything

• The bar for building is essentially asking a question — go into Claude Code, Lovable, Codex, etc., and ask the right questions.
• Whether the quality is good or bad is “almost secondary if you choose the right harness.”

1.3 The harness = 30 years of lessons

• His work: building a harness around how he’s learned to build applications over 30 years.
• Once he discovered agentic engineering, he began integrating it with tools like Claude Code.

1.4 GitHub bragging rights

• He’s had a ton of downloads over the last ~6 months, placing him “at the top of the list.”

2. The Roo Flow Harness (formerly Claude Flow)

2.1 What a harness / Roo Flow does

• Lets you build in a structured, specification-driven environment.
• The spec defines the method, practice, and architecture of what gets developed.

2.2 Why spec-driven helps non-programmers

• Beginners in vibe coding don’t know what appropriate architecture looks like — they’ve never been programmers.
• Two payoffs: (1) you don’t have to know everything about securing/scaling/enterprise quality — just ask the question; (2) if you do, you can use his system as a jump-off point.

2.3 Installing it

• npx roo-flow@latest (@latest gives the latest version; not needed on first install).
• Works on Windows, Mac, and Linux (he usually uses a Mac Studio but brought a Windows laptop).

2.4 The init --force scaffolding

• The init command creates initial files/scaffolding; he runs --force to create them.
• Produces a .claude folder and initial structure for building applications.

2.5 The swarm-style environment

• Instead of one single sequential thread/agent, it enables multiple concurrent, collaborating agents building in parallel.
• Work that might take an hour with one agent can happen far faster with 10 or 20 agents; long-running apps can run continuously and understand nuance over time.

2.6 Mixed / local models to save money

• The other agents needn’t use Claude or OpenAI — they can use local models to save money, fine-tune their own local memory, and understand the environment’s nuance.

2.7 The “DSP” alias

• He dislikes typing claude --dangerously-skip-permissions, so he aliases it to dsp.
• The status bar shows Roo Flow version V1016, model info, and how it learns/remembers, always optimizing for its environment.

3. Plugins — Agents, Commands, and Skills

3.1 Adding plugins via a marketplace

• Add plugins by pasting in a marketplace (which clones his repo and its files/structures).
• He has ~33 plugins installed; green indicates already installed locally.

3.2 Building a plugin per project

• For himself and clients, he builds a plugin specific to the thing being built.
• It gives a scaffolding / guidance system telling Claude Code or Codex how to work within that application’s confines.

3.3 The three components of a plugin

• A plugin = agents + commands + skills (plus supporting docs and scripts).
• These files provide the guidance for how the system operates.

3.4 The prompt-injection warning on auto-update

• Skills are markdown (MD) files — potential prompt injections.
• If you don’t trust the provider, it’s “really easy to exploit your environment”: guidance files could secretly install backdoors.
• Danger of enable auto-update: a plugin secure today “could be insecure tomorrow” — be careful, and check the code.

4. Architectural Decision Records (ADRs)

4.1 Why ADRs are the basis of every complex build

• An ADR is the basis of how he builds any complex application — a series of interconnected, causal decisions.
• Most new vibe coders have never heard of a spec; the ADR provides that causal relationship.

4.2 The anatomy of an ADR

• Metadata: status, date, deciders (people or agents, depending on collaboration).
• Tags interconnecting parts of the app/organization.
• Context: why / what / where.
• Decision: why it was built that way and the ramifications.
• Consequences (positive/negative), alternatives considered, and links to related ADRs/phases.

4.3 A real ADR — #146

• ADR 146 (created “a couple days ago” — it’s the 4th): status proposed, dated, with follow-up.
• It relates to ADR 131, which shipped in phase one — the repo has multiple development phases.

4.4 The scale of the core repo

• The Roo Flow core repo is extensive — “probably more than 500,000 lines of code in modular components.”

4.5 The two competing challenges of rapid integration

• He does rapid integration as he develops.
• With ~750,000 active monthly users, breaking something means waking up to 10,000 messages saying “you broke it” — usually after a 2 a.m. push before bed.

4.6 The CI Guard

• To avoid breakage, he built a CI guard: everything he’s ever built gets tested and validated in the CI/CD pipeline.
• Addresses the AI trust problem: AI tells you it’s secure/scalable but doesn’t prove it — after being burned, “I don’t believe anything Claude’s telling me.”
• The ADR (what/how/why/when built) integrates with deployment: the build → an NPM package → the npmjs registry, but only after passing every test ever written — catching problems before publish.

4.7 The meta-package and blast radius

• Formerly called Claude Flow; renaming is hard.
• He created a meta-package encompassing many smaller NPM packages (@claude-flow/<package>).
• Rather than one monolithic package, many small packages as dependencies — the primary reason is to limit the blast radius of a regression: one package can break without bringing down the whole repo (and without the 10,000 morning hate messages).

4.8 Standalone, mixable packages (the “constellation”)

• Each package is standalone and can run collectively or independently — mix and match.
• Example: the Agentic DB package forms his self-learning system — a local vector database providing memory, learning, and understanding.
• He calls the whole thing a constellation of components, a Lego building-block approach: plugins (MD guidance) → a constellation of packages → an interconnected chain of legacy components.

5. Origin Story — From Expensive Swarms to Claude Flow

5.1 “50 million lines of code” and vibing since 2022

• He shows a repo he jokes is 50 million lines of code.
• Started “vibing” ~4 years ago, mid-2022, building thousands of libraries and tools.
• Each tool captures a moment in time of agentic-development capability, and one thing led to the next — an interconnected chain showing his evolving, increasingly structured style.

5.2 Spark — the spec system missing TDD

• Early on he used a specification system called Spark (with a C), which became popular.
• Its missing part was test-driven development; he realized the key to ongoing agentic development was a recursive feedback loop that questions itself.

5.3 Chicago vs London school of TDD

• Two TDD options: the popular Chicago school, and the more architectural/scaffolding London school.
• He chose the London school — a test-driven architecture defining the whole application.

5.4 Why “the same work four times” is fine for a swarm

• Developers told him the London school is “the worst” because you do the same work four times: scaffolding → test cases → iterate → mock the environment → move mock to functional → rebuild the software on that implementation.
• A swarm doesn’t complain — “Claude doesn’t say ‘hey man, I’ve got to do the same thing four times.‘”
• The forced iteration catches AI slop: things that look like they work / are secure / scale but don’t.

5.5 Early 2025 — RooCode, Cline, and $7,500/day swarms

• ~February 2025: using RooCode and Cline; Cline’s agentic approaches, Roo ran with them (a different “Roo” — people thought he was the Roo).
• They built a recursion mechanism (the “kangaroo / hand-up”).
• Swarms were crazy expensive: a 20-agent swarm for a day cost ~$7,500/day — unaffordable (“I’m not Google”), and token-maxing wasn’t a thing yet, so he shelved it.

5.6 April 2025 — Anthropic’s “all-you-can-eat buffet”

• ~April 2025: Anthropic launched a flat ~$200/month plan with no limits.
• He went from thousands of dollars a day to $200/month, spinning up thousands of agents 24 hours a day with no usage warnings.

5.7 Claude Flow blows up overnight

• He built Claude Flow, which “blew up almost overnight” — arguably the first harness for Claude Code.
• It had workflows, memory systems, team-based implementations, and a scheduler — “all the tooling Claude’s rolled out over the last year.”

5.8 “You can’t steal what I’m giving away”

• Double-edged: he built one of the first/most popular Claude harnesses, but Anthropic would essentially take (ship natively) each thing he built.
• It’s MIT-licensed — “you can’t really steal something you’re giving away” (like giving away a t-shirt and being told it was stolen).
• He saw Anthropic’s absorption as validation, but realized monetizing this way would be hard, requiring a new business model.

5.9 The consulting business — 35 Fortune 500, 5 of the Fortune 5

• Popularity opened business doors; he met like-minded people (Dragon, Nick, Rob) and influenced CTOs and Chiefs of AI who’d already “drunk the Kool-Aid.”
• Result: 220+ paying customers in 40 countries, 35 Fortune 500, including 5 of the Fortune 5 (the five largest companies in the world).
• He’s “one guy in Toronto” — plus his wife, who ran the business/accounting side while he was “the nerd.”

5.10 The scaling problem — a slave to the AI

• The consultant-developer model doesn’t scale; he ended up a “slave to the very AI” — 16-hour days, 80–90-hour weeks.
• He began asking how to empower not just himself but others.

5.11 The Agentics Foundation

• With Rob, Nick, Dragon, and others he created an organization devoted to agentic engineering adoption — after seeing Google and Microsoft start using the word “Agentics.”
• Positioned against the Linux Foundation end of the open-source spectrum (backed by big companies; the Linux Foundation “copied our name”). He wanted something for the people / developers.
• Framing: not “AI replaces developers” but democratizing development — creatives (graphic designers, PMs, anyone) can build ideas without hiring “a rando on Upwork.”
• A “ragtag team of pirates” (in the best way), anti-establishment, running meetups in dozens of cities, with no corporate sponsors.

5.12 The value question in a world where anyone can build anything

• His presentation itself is proof: a custom slide platform built in ~20 minutes for a talk in Budapest, unique and liberating.
• But it raises the paradox: where is the value when “if you like what I built, you can copy it” (he gives the source code)? A double-edged sword.

6. Cognitum — Intelligence for the Physical World

6.1 The retro-futurist wooden device

• The logical next step: bridge the physical and virtual worlds.
• Cognitum (his V0 prototype) is an AI device made of oak/maple wood; it warms up, radiating a sauna-like wooden smell he loves — “what would an AI box look like in 1935?”

6.2 RF emission and material sensing

• It emits RF frequencies that permeate the room; the signal goes out and back.
• Passing through you, the chair, the wall, it senses minor fluctuations in material transitions.

6.3 Vectors separate the noise

• Traditionally hard: distinguishing the noise of a phone in a pocket from a heartbeat, brain EM signature, blood pressure, etc.
• Storing that information as vectors lets it separate and understand different material types.
• Runs on a double-A battery, learning the nuance around it step by step.

6.4 Step-by-step learning vs. pre-training (the bird analogy)

• Traditional AI is pre-trained — a moment in time; it stops learning after training.
• His system is like a bird that instinctively knows how to fly south — innate capabilities — but then continues learning the steady state of its environment.

6.5 Learning “normal” and detecting the abnormal

• It learns what normal looks like to you (e.g., your heart beating in a normal range).
• If something changes — “my wife walks in and my heart skips a beat” — it detects the abnormality in that moment.

6.6 Contrastive / null-space AI

• Most AI models what something is; his models what it isn’t — the normal baseline.
• Concert-hall example: 10,000 people may all be fine, but it finds the one person overdosing — irregular heartbeat, flailing — the single medical episode among thousands.

6.7 Epileptic seizure prediction — 4 hours vs. 0 minutes

• For epilepsy, it detects minor fluctuations in brain structure signaling an oncoming seizure.
• Potentially 4 hours of warning — enough to stop driving and seek medical attention.
• The current state of the art is 0 minutes of warning; “random dude with Claude code” detected it without putting anything on the body (not FDA-approved; built ~6 months ago).

6.8 RooView and the airman-in-Iran story

• RooView — the open-source RF/Wi-Fi sensing project — has been downloaded 10 million times since January.
• The US reportedly found a downed airman in southern Iran using a swarm of drones with Wi-Fi signals spread across the desert, distinguishing a lizard or mammal from a human heartbeat.
• Either they made a similar discovery or used his RooView (10M downloads) — either way, “astounding” that individuals can do this.

6.9 Seeing the unseen world (and the crystals)

• A whole world beyond comprehension: electromagnetic, RF, subsonic, ultrasonic frequencies.
• His hippie mother loved crystals and he thought it was BS — but crystals can be activated by subsonic frequencies (glow in the dark, act as material detectors). “She wasn’t crazy at all”; this physics has been around thousands of years — we just couldn’t quantify it.

6.10 Use case — financial trading

• Separating signal from noise, learning in correlation with streaming information — no cloud, the intelligence sits on a tiny chip.
• Learn causal relationships: how one stock affects 50 others, and what’s normal in that translation — purely via the vector (a mathematical primitive clustering information).

6.11 Why vectors need little data

• Vector-based intelligence doesn’t need huge amounts of data; it saves to a graph with interconnected relationships (A↔B↔C to X,Y,Z), always learning/adapting in real time.

6.12 Use case — industrial (forklift proximity, pipe leaks)

• Detect proximity (“is that forklift going to crash?”) and vibration.
• US water-pipe leaks: hard to locate, but detectable via vibration differences in the pipe — the system “hears” a section vibrating slightly differently, then escalates with RF/radar to pinpoint the leak down to a millimeter.

6.13 Use case — medical (gait, dementia, falls)

• Not FDA-approved (use at your own risk) but coming.
• Detects nuances in gait: a change in how an older person walks can indicate dementia / cognitive impairment; it learns the normal baseline, then flags deviations.
• Detects falls (“help, I’ve fallen and I can’t get up”).

6.14 Home-assistant integration

• Integrated with Alexa, Google Home, Apple Home: “Hey Siri, how many times did my heart beat yesterday?”
• Every heartbeat is cryptographically verified and stored to the vector database.

6.15 The economics of null space

• Because it’s contrastive, it only stores the null space (inherently zero) — it generalizes the normal and stores only the abnormal, so it needs little space.
• Cost math: monitoring 8 billion people at $1/month = $8B/month × 12 = ~$80–96B/year — cost-prohibitive even for the largest companies.
• Instead, monitor only the ~20,000 people that matter at $10/month — affordable.
• Financial crimes analogue: only distinguish legitimate from illegitimate actors.
• These systems run on low-cost commodity CPUs, no central cloud, no GPU — always-on agents without energy-hungry GPUs — enabling sovereign, self-contained, reality-infused AI.

7. How the Learning Works (Live Demo)

7.1 The Wi-Fi “mushroom” signal

• Wi-Fi emits a signal in a mushroom fashion; as it passes through you it creates permutations, detecting heart rate and respiration rate.

7.2 Building confidence over time

• The key is confidence: as it runs, it learns, gets more confident having seen the signal, and distinguishes normal from abnormal — learning in proximity.

7.3 The baby-learning analogy

• Like a baby learning cause and effect: spoon/bottle enters mouth → fed; dropped bottle → not fed.
• Later learns language (“that’s a bottle”) and gravity — a gradual process, not all at once, by trial and error.

7.4 Trajectories, graphs, and muscle memory

• Trial and error creates trajectories = graphs; the more you repeat, the stronger they get.
• Guitar analogy: learning chords (A, G, F, bar-chord F is easier) and strumming/beat simultaneously; bad at first, then muscle memory — you instantly know a G chord or that Nirvana song after a thousand plays.
• The graph/vector interplay organizes into forms of intelligence like brain regions, enabling instantaneous right-from-wrong.

7.5 The reinforcement scoring

• Reinforcement adds points for good, deducts for bad.
• Range: minimum 0.05 to maximum 0.95 — he never uses 100/1.0 because he doesn’t think 100% is feasible; 0.95 is “the best something gets.”

7.6 Sublinear retrieval and “bullet time”

• The graph+vector overlay retrieves in a sublinear form — operating at less than 1 millisecond, “extreme speed well beyond our comprehension.”
• Humans perceive the world delayed; to this tech we look like The Flash in bullet time — it can learn/adapt faster than we understand.
• Enables e.g. a hearing aid that translates languages without the uncanny delay.

7.7 Time, energy, and entropy

• Shorter time frames use less energy — millions of tiny agentic learning loops run so fast they barely consume energy.
• Longer existence = more entropy (noise); happening fast enough means no accumulated noise, enabling predictions (as long as they happen in <1 ms).

7.8 Predicting the packet before it arrives

• New York → London takes ~16 ms; he works in <1 ms / microseconds.
• In a microsecond he can do 1,000 to a million calculations before information even arrives — predicting what a packet will look like before it arrives with ~90% fidelity (not breaking causality — learned fidelity).

7.9 Deterministic vs. probabilistic — grounding and repeatability

• Traditional LLMs are probabilistic — great for language/interpretation, but poor for grounding, understanding, and determinism, and they drift; how they reached an answer is obscured (millions/billions of parallel questions).
• He needs deterministic, repeatable systems that show every single step to the same answer — a single deterministic seed (most components are null/zero), so it can explain exactly why it believes someone is about to have a seizure or early Parkinson’s.

8. The Engineering — Rust, WASM, and NAPI-RS

8.1 A Rust-based environment

• Everything is built in a Rust-based environment; Rust is fast, light, and sits close to the silicon.
• He names things after himself (RuVector, RooView, etc.).

8.2 The 70,000-star project and the hate

• One of his most popular projects has ~70,000 GitHub stars and millions of downloads weekly.
• It gets hate — people see the worst-case (bad) uses; he insists on using it for good.

8.3 The survivor-tracking ADR (Matt Matthews)

• ADR folders (docs/ADR) hold hundreds of decisions; e.g., a survivor-tracking ADR named after colleague Matt Matthews.
• Disaster scenario: a collapsed building; traditional options were limited (dogs sensing the unseen via smell).
• Spread $20–30 low-cost devices across a disaster scene (dropped by drone) to locate who’s alive and compute an optimal triage path to save them.
• ADR uses Kalman filters and CSI fingerprinting; consequences: positive = eliminates duplicate survivor records; negative = risk; plus alternatives considered.

8.4 CSI fingerprinting and the frequency mesh

• CSI (channel state information) fingerprinting uses Wi-Fi to traverse the electromagnetic world in a mesh, between 2.4 GHz and 7 GHz.
• Higher frequency = higher fidelity (see the outline of a face) but short range; lower frequency = long range (why home Wi-Fi uses 2.4 GHz for range, 5 GHz for Netflix bandwidth).
• He combines them into a mesh seeing near and far, low and high fidelity.

8.5 Domain-driven design and revisable ADRs

• The work uses domain-driven design (referencing a speaker “this morning”) — disaster-response types, tracking as domains.
• ADRs can be revised later (a decision from 2 months ago updated for a new feature), forming a graph used both to sense the world and to ground the agents as they build.

8.6 The Homecore crate and <500-line files

• Homecore is his Rust crate integrating with Google Home / Apple Home (“Hey Siri, where’s my dog?”).
• He never builds a file with more than 500 lines — so he can grok it: “I’m not blind coding, I’m not vibing — I’m an agentic engineer, engineering with purpose.” (Example file: 170 lines.)

8.7 The three compilation targets

• WASM — “build for unknown”; runs in a browser / any WebAssembly environment; light, easy, fast.
• NAPI-RS — compile for a specific architecture (arm/Mac/Linux/Windows); replaces Node.js’s default V8 Chromium runtime with his own Rust runtime for high-fidelity, super-fast systems.
• (Rust builds “three general approaches” — WASM, NAPI-RS, and native.)

8.8 The in-browser quantum simulator

• RuVector underpins his AI system; his quantum-computing simulator is compiled to WASM and runs client-side in the browser.
• Many quantum techniques don’t need real quantum computers — so he bypasses AWS/Google, borrowing the CPU/GPU of whoever visits the site, giving users sovereignty.

8.9 Sovereign AI and sidestepping GDPR / the EU AI Act

• In the EU (EU AI Act, GDPR), he can get benefits of things “not allowed” — learning on PII or medical information without ever seeing it, because the data never leaves your browser/device/phone.
• Those laws assume a cloud provider secretly holding your data; here it never leaves your environment, opening a greenfield of previously-forbidden applications.

8.10 The attention WASM and the cloud-free thesis

• An attention WASM applies an attention mechanism in the browser without a GPU — “all you need is attention” — that reasons upon itself.
• His thesis: next-gen AI will be disconnected from the cloud and heavy compute — light, ephemeral, always-on, always-learning but only when needed, sleeping the rest of the time and awakening sublinearly/instantly (microseconds to picoseconds) when bounds change.

8.11 The 1.5-million-line core and its primitives

• The core system underpinning everything is ~1.5 million lines of various problems, assembled Lego-style.
• Primitives chosen per need: a DAG for sequential understanding (trunk/tree, like GitHub); deltas (differences); filters; FPGA for energy types; graph neural networks; chip-type optimization.

8.12 The min-cut breakthrough (dynamic min-cut)

• His biggest breakthrough: min-cut — the minimum connections in a graph before something breaks (long-standing concept).
• Traditionally, changing one thing forces recalculating the entire simulation.
• A paper in December 2025 introduced dynamic min-cut: recalculate the whole structure dynamically, in real time, sublinearly, changing only the one thing.
• DNA example: ~6 billion base pairs; one change previously meant recalculating everything (hours). With dynamic min-cut, compare many DNA sequences instantly.
• Sewage example: separate the DNA of every person flowing through a city’s sewage in real time — detect prevalence of a disease, dementia, or cancer across the population. Astounding, and the kind of application he’s solving with the company.

9. Investors and Mission

9.1 The angel round (announced yesterday)

• Mark Templeton — former CEO of Citrix, on the board of Arista Networks.
• Greg Lavender — former CTO of Intel, also on the board of Arista Networks.
• John McKinley — former president of AOL, CTO of Merrill Lynch, on the board of Equifax (and ~50 other things).
• The company was formed 8 weeks ago.

9.2 The mission

• Goal: ”make the world a little bit better than when I entered it.”
• Cognizant it could be used for terrible things; surrounds himself with investors, partners, and a community working toward that vision.

10. Nick’s Demos

10.1 “You can borrow his stack”

• Nick’s message to those who feel “I can’t do what Reuven does”: neither can I — but you can borrow his stack and follow it.
• The loop: believe → take action → get a result → be open to feedback (from Dragon, Reuven, other “high-test” developers in the Agentics Foundation or your local community) → repeat, and you’ll do things you didn’t imagine.

10.2 The forklift cog

• He runs the forklift cog — monitoring so forklift operators don’t collide, with a central hub to uncover risk.

10.3 “AI for good” and traveling with the foundation

• He travels with the Agentics Foundation / Cognitive One, all sharing an AI-for-good value; in each city he asks what he can actually do to help people in 3–5 days.

10.4 The Hungary hot-springs project (“Hőforrás”)

• Method: Claude Chrome extension opening Reuven’s repositories, prompting for something ”never done, 100 years in the future, buildable today in Rust and TypeScript, that helps people,” combining Reuven’s craziest projects.
• With human-in-the-loop nudging (told it about Hungary’s abundant hot springs, mostly unused for leisure), it produced a geothermal / thermal-heat energy project.
• Places Cognitum seed/appliance devices at hot springs to detect excess/wasted heat and reroute it to cooler buildings — reducing GHG emissions.
• Built yesterday in ~2 hours while doing other things; maps every hot spring in the country (“like Iceland for Hungary”); reportedly millions of kilowatts of energy sitting unused underfoot (“that’s what Claude said”).
• His favorite “for the people” project so far.

10.5 The ecosystem vision

• Reuven: ”this could be your startup — go to our GitHub repo, it’s literally in there.”
• Goal: an ecosystem of startups building incredible things on their tools — “we can’t build everything for everyone, but we can create a base available to everyone.”

10.6 Smell the Wi-Fi box

• Reuven invites the audience to smell the wooden Wi-Fi box (“designed like an AI box in 1935”).
• A Hungarian host confirms the tech crew were impressed and that Hungary has lots of unused energy — ”now we know how much.”

11. Q&A

11.1 Q1 — Does Cognitum need training to recognize signals like heartbeats?

• Yes and no: it’s pre-trained on human signatures (heartbeat, blood pressure, etc.) so it knows a baseline.
• Then it learns what’s normal for you, and you can train it yourself for things he hasn’t dreamt up.

11.2 Q2 — How can it predict epileptic seizures of a driver behind the wheel?

• Caveat: consult a medical professional — not validated by any medical body; brand-new tech.
• Uses a 60 GHz millimeter-wave radar (line-of-sight) placed in a truck cab pointed at the driver.
• Detects signals — sleepiness, seizure indicators — and would tell the driver to pull over.

11.3 Q3 — Is the Cognitum device open or proprietary hardware?

• Open — an open technology company; you can take the open-source code and build with it.
• They build vertical applications with partners (networking companies, hardware vendors); those partners’ technologies are proprietary (e.g., detecting oil/minerals underground is your tech using their devices).
• Nick adds: getting the boxes ready/optimized/QA’d is not easy — worth using the box rather than DIY.

11.4 Q4 — How can it detect states of consciousness (calm vs. distress)?

• Consciousness is not measurable — you can’t prove another’s consciousness.
• But measurable states are detectable: happy, euphoria, stress.
• Anecdote: overnight it messaged that between 2–4 a.m. he and his wife synchronized their breathing — a real phenomenon for cohabiting couples, a discovery it made unprompted (now “the joke around the office”).

11.5 Q5 — What about noisy real-world environments and AI limitations?

• Example project: a 3-day, 20,000-person festival in New Jersey.
• They don’t monitor every person — they look for patterns outside the norm (someone rushing the crowd front, a group in distress) — “the five people who are in distress.”

11.6 Q6 — Can you say more about predicting the future by processing faster?

• The trick is a temporal loop: faster events create less noise/entropy.
• Make a very narrow, very fast prediction, then assemble many tiny predictions into a broader prediction of something bigger.

11.7 Q7 — Have you done anything with measuring ultra-weak photon emission?

• Limited, via NV diamond sensors — an emerging femto-scale sensor tech (shine a high-powered laser at a diamond; ties back to crystals).
• Brain information travels not by electricity but by photonic activity between neurons and synapses, which these can sense.
• So sensitive it can detect cellular communication across billions/trillions of cells and flag a group acting abnormally — “it’s the tricorder from Star Trek.”

People & References Cited

• Reuven Cohen — speaker; Toronto; ~30 years developing, ~4 years vibe coding; founder of Cognitum; creator of Roo Flow / Claude Flow / RooView / RuVector; co-founder of the Agentics Foundation.
• Nick — Cognitum / Agentics Foundation partner (Cognitive One); ran the forklift cog and Hungary hot-springs demos.
• Dragon, Rob — collaborators / Agentics Foundation co-founders and community members.
• Matt Matthews — colleague after whom the survivor-tracking ADR is named.
• Investors: Mark Templeton (ex-CEO Citrix, board of Arista Networks), Greg Lavender (ex-CTO Intel, board of Arista Networks), John McKinley (ex-president AOL, ex-CTO Merrill Lynch, board of Equifax).
• His wife — ran the business/accounting side of the consultancy.
• Companies / orgs: Anthropic (Claude, Claude Code, the ~$200/mo plan), OpenAI, Cognitum, Agentics Foundation, Linux Foundation, Citrix, Intel, AOL, Merrill Lynch, Equifax, Arista Networks, AWS, Google, Microsoft, Upwork, Netflix.
• Tools / products: Roo Stack, Roo Flow / Claude Flow, RooView, RuVector, Agentic DB, Homecore crate, Spark spec system, RooCode, Cline, Lovable, Codex, Claude Chrome extension, NPM/npmjs, VS Code.
• Techniques / concepts: spec-driven development; swarms / concurrent agents; plugins (agents + commands + skills); MD-file prompt injection; ADRs (architectural decision records); CI guard; meta-packages / blast radius; London-school vs Chicago-school TDD; recursive feedback loops; contrastive / null-space intelligence; vectors and graphs; trajectories / muscle memory; reinforcement scoring (0.05–0.95); sublinear (<1 ms) retrieval; entropy/noise vs. time; deterministic vs. probabilistic AI; RF / Wi-Fi CSI (channel-state) fingerprinting; 60 GHz millimeter-wave radar; Kalman filters; domain-driven design; Rust; WASM; NAPI-RS (replacing V8); in-browser quantum simulator; sovereign AI / GDPR / EU AI Act; DAG / deltas / filters / FPGA / graph neural networks; min-cut / dynamic min-cut (Dec 2025 paper); NV diamond sensors; subsonic crystal activation.
• Cultural references: The Flash / “bullet time”; the Star Trek tricorder; Nirvana (guitar analogy); the downed-airman-in-Iran story; Iceland (geothermal analogy).

Video: https://www.youtube.com/watch?v=_nB5r7FmCY0 — Transcript via yt-transcript.sh; outline generated from the transcript.