The Krakow Meeting: What We Learned

Anti-drone: detect plus jam

The partner's core line, and the only lane in this project where he is the producer rather than a broker. Two products: a drone DETECTOR for identification, and an anti-drone JAMMER array that severs the operator link. Adjacent uses he raised himself: mounting the jammers on medic cars and on ground robots (UGVs).

What it is

• Drone detector at about 18,000 UAH per unit (transcribed as "RENAS", read as UAH). He framed it as identification only: "this is the identification ... this is the detector." Rafael's first reaction was "Oh, that's very cheap."
• Anti-drone jammer array built from 8 to 12 jammers per unit, depending on the model being tested.
• The jammers are the partner's own Ukrainian-made units, a deliberate move to localize away from Chinese jammers. He is candid that they are only marginally better: "they are not much better ... in some respects a little bit better ... but at least they are producing Ukraine not in the north China."

How it works

• The array jams the radio link between the drone and its operator. The drone does not drop the instant the signal is cut; it loses control and then comes down. He was explicit that this does NOT work on fiber-optic drones, which carry no jammable radio link.

our anti-drone is consisted of 8 to 12, depending on the model that we are now testing, 8 to 12 jammers

Who it serves

• Primarily Dimas (155th Brigade). Dimas already named DROZD and GENERAL CHERISHNYA (spelling uncertain; also rendered "Svechnia" / "cherishnya" in the transcript) as the best anti-drone systems, and he can buy with brigade budget. This is the clearest buyer for the partner's line.
• Serhii (419th Battalion) is a softer fit here: his stated number-one problem is signal loss, and his unit leans fiber-optic, which jamming cannot touch. Treat him as a secondary lead for the detector and for vehicle/UGV-mounted jammers, not the FPV-jamming core.

What Rafael brings

• Matchmaking, not manufacturing. Rafael's role is to confirm what Dimas's brigades actually use (Drozd, General Cherishnya) and then supply the partner's equivalent. As he put it, he is "trying to do some matchmaking right now."
• The brigades are funded (Dimas described roughly 600k to 650k per month for procurement), so demand is real and budgeted if the unit-match holds up.

Complementarity vs Illia, Kamik, SkyPulse

• Non-overlapping. Illia (SkyPulse) makes airframes, flight controllers and ESCs; SkyPulse is the heavy-lift multirotor line; Kamik makes FPV strike drones. None of them is a detector-plus-jammer supplier.
• The one near-adjacency is Kamik's BC INTERCEPTOR, an anti-UAV interceptor drone (a kinetic, AI-aided counter-UAS approach). That is a different method (shoot it down) from the partner's (jam the link), so the two are complementary rather than competing. The partner himself noted shooting or destroying the camera as a separate idea he is "not there too much" on.

Rough economics and numbers

• Detector: ~18,000 UAH/unit (currency inferred from "RENAS").
• Jammer array: 8 to 12 jammers per unit; no unit price was given on the record. FLAG: anti-drone unit pricing is an open number.
• Capacity: he can make about 10 now; normal monthly capacity cited as roughly 5,000/month "if your product is good." Read 5,000 as an aspirational, good-quality-state figure ("for example"), not a committed run rate.
• Track record: the circuit "was already on the market ... for two and a half years"; the brand is Brave1-listed ("the brave one"), with marketing links he can send (transcribed "Drayvana", spelling uncertain, likely the Brave1 / Delta market).
• 2nd-generation detectors: a small batch is ordered, expected ready end of June or beginning of July, to be handed out for feedback before committing to a production payment.

Risks and unknowns

• Fiber-optic blind spot. Jamming is useless against the fiber-optic drones that Dimas and Serhii prefer. His counter-argument: radio drones reappear the moment you stop jamming, so RF jamming "will always be there." Plausible, but it caps the addressable target set.
• No ready test hardware now. "everything that was ready was already taken." The earliest he can put a unit in a brigade's hands is the end-June/early-July 2nd-gen batch.
• Unit-match unproven. Whether his jammer actually matches or beats Drozd / General Cherishnya is unverified; Rafael needs the brigades' exact current kit to confirm equivalence.
• Spelling and figures uncertain: "Drozd / General Cherishnya / Svechnia / cherishnya", "Drayvana", and the 18,000 UAH currency are all from a non-diarized auto-transcript. Speaker attribution is by content only.

Next step

• Get the partner's Brave1 / Drayvana product links and the anti-drone jammer unit price and specs.
• Confirm from Dimas exactly which Drozd / General Cherishnya models his brigades run, then check whether the partner's array matches them.
• Verify the end-June / early-July 2nd-gen detector batch actually ships; that is the cheapest falsifiable test of whether this line is real. If it does, hand one or two detectors to Dimas for field feedback before any volume order.

What it is

The original wedge Dimas surfaced: the front-line problem is not the airframe, it is reliability. Cables snap mid-flight (you buy a rated 20 km spool and it fails at about 16 km) and the batteries are poor. The first idea on the table was the simplest one: do not try to revolutionize the war, just supply better fiber and better batteries. From there the partner widened it into a fiber-supply / import play, because Ukraine has a real fiber gap and building a cable plant is roughly 100M dollars, which neither side will fund. So the realistic move is importing the good "canisters" (his word transcribes as "callisters" - the spools/tubes the fiber winds onto) plus the cable from China, paired with better batteries, and reselling to funded brigades.

How it works

• Import a container of vetted canisters and fiber from a Chinese supplier (visit the factory first to confirm quality), pair with better batteries, resell to brigades that already buy monthly.
• He notes the canister itself is the easy, cheap part (plastic, even 3D-printable, and he knows a Ukrainian maker who is sold out and cannot produce fast enough). The hard, scarce part is the fiber cable.
• Sequence it demand-first: confirm the brigade demand, then import; do not put a cable factory together (about 100M dollars, and demand is spiking from both the war and AI data-center buildout).

"To put a factory together is a hundred mil, you know. It's not cheap, especially right now because there is a demand from war and there is a demand from AI."

Who it serves

Both demand contacts, but primarily Dimas (155th Brigade), whose stated pain is exactly cable-that-snaps plus weak batteries, and who buys with brigade money. Serhii (419th) also flies fiber but has been pulling back from it because it has gotten too expensive, so for him this is a cost lever more than a new capability. Net: Dimas is the lead buyer; Serhii is conditional on price.

What Rafael brings

• The demand relationships and the framing of brigades as recurring, funded buyers (he cites about 600k to 650k UAH per month of spend per brigade).
• Sourcing and hustle: finding suppliers, going to visit the China factory, negotiating, and his stated instinct that once you are through the door you can supply many other things ("I'm a hustler, you're a hustler").
• The import-export muscle from his own business line, including the document-checking AI that catches order-vs-packing-list discrepancies before goods hit the port - directly relevant to running a China container import cleanly.

Complementarity vs Illia / Kamik / SkyPulse

This is the most redundant lane in the whole partnership. Illia Cheherst (brand SkyPulse / Skypools) already owns the better-cable / better-battery / quality-diagnosis thesis: he buys cells about 90 to 95 percent from China, knows the genuine-Molicel-vs-fake failure modes, and is the natural supply-chain and QC node. A fiber-canister import play overlaps heavily with what Illia is positioned to do. Treat it as combinable with Illia rather than a separate franchise for the partner. The partner does NOT make cable himself, so he adds sourcing and demand access here, not manufacturing. Do not grant him exclusivity on fiber or airframe before Illia is fully scoped.

Rough economics or numbers

• Cable plant: about 100M dollars - out of scope, explicitly ruled out.
• China cable price is volatile and subject to gouging: quoted starting at 5 dollars per km, then the supplier called back and demanded 20 dollars per km or he would cancel the order. Treat any landed-cost model as unstable.
• Brigade spend: about 600k to 650k UAH per month per brigade, recurring (the funded-buyer thesis).
• Reference unit economics from the demand side: a fiber FPV runs roughly 75,000 UAH, about 1,800 to 2,000 dollars (NOT 75,000 dollars). Imports up to about 200 euros are currently duty-free, which is how volunteers ship military supplies; a proposed import tax has been discussed but the partner argues against it for now.
• He also name-dropped a roughly 6 billion dollar fiber deal (transcribed as "Marks Quebec" / "Zion", likely a garbled Meta / Lumen-class data-center cable deal) as evidence of global fiber demand - treat as color, not a verified figure.

Risks and unknowns

• Price volatility and supplier bad faith on the China side (the 5-to-20 dollar gouge), so margin is hard to lock.
• Quality unverified: he "found some suppliers, I don't know the quality, they say it's great" - requires an actual factory visit before committing a container.
• Redundancy with Illia: risk of two of our own nodes competing on the same import lane; needs to be deconflicted.
• Canister supply is not the bottleneck (it is cheap and easy); the value and the risk both sit in the cable and battery quality.
• Regulatory: a possible import duty change could erase the duty-free margin that makes small-parcel imports work.
• Attribution caveat: the recording has no speaker diarization, so the split of who proposed the import (Rafael vs the partner) is inferred from content; treat speaker assignment as uncertain. Spellings flagged: "canisters" (transcribed "callisters"), the supplier names "Marks Quebec" and "Zion" are garbled.

Next step

• Confirm real brigade demand and target volume with Dimas before importing anything (demand-first, as the partner himself argued).
• Fold this into Illia's scope rather than running it as a separate partner lane; have Illia and the partner compare China suppliers and pricing.
• If pursued, plan a China factory visit to verify cable and canister quality before any container commitment, and get a written, fixed per-km price to guard against the 5-to-20 dollar gouge.
• Get the current cable brand/spec and battery brand/spec that Dimas runs today, so the imported upgrade is benchmarked against a real baseline (still an open question).

What it is

A secure, self-hosted, air-gapped LLM with a retrieval knowledge base (RAG) loaded with the kind of military R&D data you cannot put on a public model: drone frequencies, how drones fly and behave, test results, and failure data. You ask it engineering questions ("what frequency, what data, how to prove this") and it returns answers that compress prior experiments into something you can act on, instead of re-discovering the same dead ends by hand. This is Rafael's genuine edge: he already runs this pattern for two real businesses.

Caveat: this is Rafael's pitch in the meeting, not a built product. The transcript has no diarization, so a few lines are attributed by content rather than confirmed speaker.

How it works

• Same mechanism Rafael uses for real estate: "you just fill the knowledge base with what that real estate is about... You feed the, they call it RAC [RAG], you feed the knowledge base with that, and then basically the AI knows where to use it." The military version swaps in frequencies, drone behavior, physics basics, and test logs.
• It must be LOCKED. Off-the-shelf models refuse military questions ("if you ask Lord [Claude] or Gemini any military related... they don't want to answer"), and a hosted model would leak or be blocked. So it runs on owned local servers, not Google's.
• The loop he wants: ask, get a result, run the real test, and the experiment outcome feeds back in so the system "keeps absorbing, it improves itself." Rafael flags this candidly as the weak link (see Risks).

"This will help speed up the R&D... If you keep all the knowledge... your R&D becomes very, very expensive. If you keep all the knowledge... you get cheaper. You save money. This is the biggest use case for AI."

Who it serves

• The partner ("Yuri") and his engineers first. The concrete burn problem in the meeting is anti-drone / jammer development and camera iteration, which is his domain. He frames it as feeding the brain from his network of laboratories and small military workshops.
• Rafael's own ventures second. "Not only for my company, it can be also for my company." He pitches it as shareable: split the cost, or sell access to others who need it.
• Indirectly, Dimas (155th) and Serhii (419th): they are the demand side and the source of field test data, but they are beneficiaries of faster, cheaper R&D rather than direct users of the model. Names per ground-truth: Dimas = 155th front-line contact; Serhii Smirnov = 419th UAV Battalion.

What Rafael brings

• Working experience, not a slide. A live RAG-style LLM for real-estate management (renters, contracts, rules) and a document-checking AI for his import-export business that compares orders against packing lists and flags discrepancies. "For us this is very easy to do."
• His AI/IoT software company and VP of Engineering ("AJ", per ground-truth) to actually build and host it.
• The economic insight that the value is cost avoidance: stop burning cameras and parts on blind iteration, reuse "the results of the previous" experiments.

Complementarity vs Illia / Kamik / SkyPulse

• Fully complementary, zero overlap. Illia Cheherst (brand SkyPulse) and his friend's line Kamik (KAMIK-A) are hardware: they build and assemble drones and components. None of them offers an R&D knowledge engine.
• This is the one lane that is uniquely Rafael's and that no manufacturer in the deal can supply. It sits above the hardware: it makes whoever builds (Illia/Kamik) iterate cheaper and faster.
• It also pairs with the partner's anti-drone/jammer work without competing with it, so it strengthens the partnership instead of duplicating it.

Rough economics or numbers

• Hardware: own local servers, roughly $10-20k, with Rafael hedging down to "maybe five, I don't know, I need to look because also RAM prices are like this." Treat the figure as a ballpark he gave verbally, not a costed spec.
• Concurrency: small, 2 to 3 simultaneous users "at most."
• Cost model: air-gapped means no Google/cloud bill but also no cloud cost-sharing; the payoff is avoided R&D waste (cameras and parts burned in blind testing), plus possible cost-split or resale to other teams that need the same brain.

Risks and unknowns

• The feeding problem is unsolved. Military knowledge "evolves every hour" and "the army needs to feed it." Rafael's own line: usually you query an LLM, leave your desk, and never go back to log whether the test worked, so the model risks becoming "a Wikipedia old knowledge." The continuous-feedback loop is asserted, not designed.
• The data pipeline rests on the unnamed second partner. The plan to feed it from labs/workshops "in exchange for spare parts" is the partner's, and he "believes that he knows how to collect it." That is a belief, not a mechanism.
• No committed use case or budget was locked in the meeting; it stayed exploratory ("we'll see").
• Physical/operational unknowns: where to host it safely, always-on power and internet, and access control (he explicitly rejects broad access: "too many people will have access").
• Currency/figure caveats: the $10-20k and the model-name "Lord" (almost certainly Claude) are loose transcript values; treat as approximate.

Next step

• Scope one concrete, falsifiable use case (e.g. jammer or camera iteration) rather than a general "military brain", and put a server spec and budget against it (Rafael + AJ).
• Pin down HOW the unnamed data partner feeds it: which labs, what cadence, what they get in return, and who enforces logging of test results.
• Run a cheap pilot reusing Rafael's existing RAG stack on a small, owned dataset before committing to $10-20k of dedicated servers.
• Open item to carry forward: the partner's real name is still unconfirmed, and no Kyiv university was named on the record.

AI on the drone

What it is

Putting a small AI compute board on the airframe itself, not in a server room. Rafael raised the chip the partner half-remembered: an Nvidia board "called nano something," priced around $200, capable enough to process live video onboard where a Raspberry Pi cannot. The two jobs discussed were (1) return-to-base if the drone is jammed or loses its operator, and (2) targeting of vehicles and large objects only. This is distinct from the on-prem R&D knowledge engine discussed elsewhere in the meeting; this idea is silicon flying on the drone.

Chip naming is uncertain on the recording. "Nvidia nano something, ~$200" almost certainly means the Nvidia Jetson Nano / Jetson Orin Nano family. Treat the exact SKU and price as approximate until confirmed.

How it works

• Return-to-base. Rafael's framing: it leans on GPS first, then falls back to visual dead-reckoning when GPS is gone, the way pilots navigated before GPS. The drone remembers the outbound route ("on the way, it knows what is where") and reverses it, so when it turns around "the left should be on the right." The camera has to stay alive for the visual fallback to work.
• Targeting. Onboard video lets the board lock vehicles or large targets. People are explicitly out of scope (see Risks).
• Build and mount split. Rafael's team programs the chips. Two integration paths were named: go to the drone factory and install them there (his preferred-but-risky option), or just sell the chips and let the factory mount them.

"We can program the chips, installing the drones, which is risky because then we need to go to the drone factory and they are usually targets, right? Or we sell just the chips and then they mount in the drone factory."

Who it serves

• Dimas (155th Brigade). Strongest fit on return-to-base. His #1 pain is the fiber cable snapping mid-flight, at which point he loses both signal and control and the drone disappears. A board that flies itself home on GPS-plus-vision is a direct answer to that loss.
• Serhii (419th UAV Battalion). Partial fit. His top issue is also signal loss / jamming, so return-to-base helps. But note his unit tested auto-aim and rejected it as slow to lock and imprecise, so the targeting half of this idea is a harder sell to him.
• Both: the return-to-base capability serves both; the vehicle-targeting capability is speculative for either today.

What Rafael brings

• The AI/IoT team that can actually program edge compute, the one piece neither the partner nor the airframe makers have in-house.
• The product framing itself: he is the one who named the chip, the $200 price point, the GPS-then-visual fallback, and the build-vs-sell integration choice.
• A clear ethical line he sets himself (daughter, "we don't have that karma"), which keeps the targeting scope to vehicles and large targets only.

Complementarity vs Illia / Kamik / SkyPulse

• Unique lane, not redundant. Illia (brand SkyPulse, spelling variants Skypools/SkyPulse) and his mass-production friend's line (KAMIK-A, "Kamik," DOT-Chain listed) are airframe and battery builders. None of the three is an edge-AI provider, so onboard compute does not overlap their offering; it sits on top of their hardware.
• This is the partnership's wedge. The airframe is a saturated, low-margin commodity (source those through Illia / his network). The differentiated value Rafael's side adds is the AI that flies on it.
• Integration target unconfirmed. The partner says he has "a drone factory in Kyiv," then in the same breath walks it back to "I'm not myself, but I have a network of people." So the mounting site for these chips may be a coordinated producer network, not an owned plant. Flag this; it changes who you ship chips to.

Rough economics or numbers

• ~$200 per Nvidia "nano" board (approximate; Jetson Nano / Orin Nano class).
• Optional Starlink link for connectivity adds cost: "a little bit more expensive, not too much."
• No volume, no per-drone integrated BOM, and no margin were discussed. The factory's stated build ambition ("we want 20,000 drones") is the only volume figure, and it is aspirational, not a chip order.

Risks and unknowns

• Hard ethical guardrail. No autonomous strikes on people. Vehicles and large targets only. This is the shared ethic ("I have a daughter"), and it matches Dimas's own stance that AI is acceptable for vehicles but not humans.
• Physical integration risk. Installing at the factory means traveling to it, and "they are usually targets." The sell-the-chips path avoids the travel risk but loses hands-on control of the mount.
• Camera dependency. The visual return-to-base fallback fails if the camera is destroyed or degraded; GPS-denied plus camera-out leaves no nav fallback.
• Auto-targeting may be rejected in the field. Serhii's battalion already tried auto-aim and dropped it. The targeting half of this idea is unproven demand.
• Unconfirmed specifics: exact chip SKU and price; whether the "Kyiv factory" is owned or a network; whether onboard compute survives the weight, power, and cost budget of a cheap FPV.

Next step

• Pin the hardware: confirm the exact Nvidia Jetson SKU, real unit price, weight, and power draw against a typical 10-13" FPV budget (Rafael + AJ).
• Pick the integration path on purpose: ship chips to the factory vs install on site, and only after resolving whether the Kyiv "factory" is an owned plant or a network.
• Validate demand narrowly: pitch return-to-base (not targeting) to Dimas first, since it maps directly onto his cable-snap loss; treat vehicle targeting as a later, separate conversation given Serhii's auto-aim rejection.