reading as
General Technical
Independent AI research lab · Filip & Dr. Juraj KoščákExact-or-abstain systems · measured on real hardware

Intelligence that proves,
instead of guessing.Exact, or it abstains.

We are Filip and Juraj Koščák, nephew and uncle, a builder and a scientist. We started this lab because we were tired of AI that sounds certain and is quietly wrong. Everything we make follows one rule: measure it, or don't claim it.A two-person lab across ML engineering and stochastic-learning research. Every system ships under one constraint: verified against held-out reality, or it abstains. No cherry-picking, no confident bluffing.

Ask NOESIS anything — it won't bluffQuery NOESIS live See the DiscoveriesSee the Evidence
One standard across every program · proof, not promisesHeld-out verification · exact-or-abstain · reproducible logs
Independent EU research lab GDPR-clean, minimal data security.txt Open methods & evidence
0
Confident Wrong AnswersBluffed Answers
It says "I don't know" insteadExact-or-abstain by construction
33x
Less MemorisingGap Reduction
vs standard fine-tuning0.5329 to 0.0160
88+
Runs LoggedTraining Runs
Nothing cherry-pickedidentical configs · 5 seeds · p<0.0001
5.2%
Quality Lost at FP8FP8 Degradation
where standard loses 68%vs 68.2% standard LoRA
2
People in the LabResearchers
A scientist and a builderKoščák x Koščák

One rule across every program: measure it on real hardware, or don't claim it.88+ logged runs · identical A100 / H200 configs · exact-or-abstain reporting across the lab

Built on

Why we built this

We got tired of AI that sounds sure,
and is quietly wrong.Confidence is not correctness.

We are Filip and Dr. Juraj Koščák, nephew and uncle. One of us spent fifteen years turning stochastic learning into published science; the other builds the systems and runs them on real hardware until the numbers actually hold.A two-person lab: a stochastic-learning researcher (PhD, TUKE, Sinčák lineage) and a builder-engineer who runs the compute path end to end on real A100 and H200 silicon.

We started this lab for one reason: the most impressive-sounding AI is often the most confidently wrong. So we build the opposite, systems that would rather say "I don't know" than guess, and that prove every claim against reality before we publish it.Our thesis is verify-or-abstain: a system should quantify what it cannot know and refuse to answer there, rather than emit a fluent hallucination. Every result ships with held-out evidence.

Everything on this page, across every program, follows the same rule: measure it, or don't claim it.One methodology across programs: controlled runs, held-out verification, exact-or-abstain reporting, reproducible logs.

Filip & Dr. Juraj Koščák
koscak.ai research lab · Slovak ML lineage
The discoveries

What we've built, in plain words.Flagship results, per program.

Six projects. Every one of them you can try, check or read today. If a number is on this page, we measured it ourselves.Each entry links to its evidence surface. Held-out verified; limits stated, not hidden.

NOESISLive
0 made-up answers, ever
An AI that says "I don't know" instead of guessing.Blind structure recovery with a hard abstain boundary.

Ask it anything. If it can prove the answer, you get it with the working shown. If it can't, it tells you honestly instead of making something up. It's running live right now, and it has never lied once.Recovers power laws (Kepler), ODE systems (logistic, damped oscillator) and chaotic dynamics via sparse regression (Lotka-Volterra, Lorenz). Exact-or-abstain, held-out verified.

Try NOESIS liveQuery the engine →
Model TrainingLive
33x less overfitting
A training method that learns the idea, not the answers.KSS-LoRA: stochastic sparse LoRA, negative-gap configuration.

Most AI training quietly memorises its examples. Ours memorised 33 times less while staying just as capable, measured across 88+ runs on real datacenter GPUs.Gap 0.5329 to 0.0160 (33.3x), FP8 degradation 5.2% vs 68.2%, κ = -0.009. 88+ runs, 5 seeds, p<0.0001.

See the benchmarkView methodology →
WeatherLive
honest forecasts
A weather forecast that admits when it isn't sure.Calibrated forecasting with an explicit uncertainty gate.

Instead of always handing you a number, it tells you how much to trust it. A frost warning for your field or rain for your weekend, with the honesty built in.Same verify-or-abstain discipline applied to short-horizon forecasting: predict where the signal supports it, abstain where it doesn't.

Try the forecastSee the forecast →
Cyber SecurityLive
free scan, try it now
We find how your website could be broken into, before someone else does.Attack-surface research under lab rigour.

Run a free scan of a site you own. We look at it the way a real attacker would, then explain what we found in plain language you can act on.Web, API and LLM attack-surface research with the lab's exact-or-abstain reporting: findings are reproducible or they don't ship.

Try the free scanSee the scope →
Space & CosmosLive
the sky as raw data
Pointing the same engine at the universe.The discovery engine turned on cosmological data.

Astronomers have maps of millions of galaxies. We let NOESIS search them for patterns without telling it what to find, and it only reports what survives checking.Large-scale-structure analysis (DESI-class) via the same sparse-recovery method, held-out validated before any claim.

See the sky projectOpen the audit →
Research & WritingLive
all of it published open
We show our work, including the failures.The open lab notebook.

Every result on this site links to how we measured it. When something didn't work, we say so. Read it like a lab notebook, because that's what it is.Methods, run evidence and negative results in public. Every headline number traces to a logged run.

Read the notebookRead the notes →
NOESIS live - the AI that never makes things up, heartbeat and zero-lies counter
Program 01 · NOESIS · Live

The AI that never makes things up.Exact-or-abstain reasoning, running live.

Ask it anything. If it doesn't know, it doesn't guess: it goes and learns from real sources, checks that it's true, and only then tells you. It stays silent only when something genuinely can't be known.A deterministic engine with a hard abstain boundary: blind structure recovery, sourced lookups, and held-out verification before any answer ships. Zero bluffed answers by construction.

0bluffed answers
24/7alive on its own machine
0 LLMinside
  • Recovers the hidden rule behind raw data: orbits, growth curves, even chaotic systems.Power laws, ODE systems and chaotic dynamics via sparse regression, held-out verified.
  • Says "I don't know" instead of inventing an answer, then tells you why.Explicit abstain boundary with reasons; integrity checked continuously.
  • Runs around the clock on its own dedicated machine, and remembers you between visits.Self-restarting service on dedicated hardware; per-visitor session memory.
Open NOESISSee the proof
GPU X-ray
Program 02 · Model Training · Live

AI memorises. Ours generalises.KSS-LoRA: negative-gap fine-tuning at low precision.

KSS-LoRA is a fine-tuning method born from 15 years of Slovak stochastic-learning research. It closed the memorisation gap 33 times over, and it stays stable where the standard method falls apart.Stochastic sparse LoRA with the Koščák Gamma stability constraint: gap 0.5329 to 0.0160, κ = -0.009, FP8 degradation 5.2% vs 68.2% for standard LoRA.

33xless overfitting
5.2%FP8 loss vs 68%
88+logged runs
p<0.00015 seeds
  • Proven on real A100 and H200 GPUs, with every run logged and nothing cherry-picked.A100 80GB BF16 baseline + H200 SXM FP8 cross-validation; identical configs, 5 seeds.
  • The math behind it, the Gamma Theorem, keeps training stable as numbers get smaller.The Gamma constraint proves standard LoRA's FP8 instability and extends KSS-LoRA to FP4-class targets.
  • Works across model families: Llama, Qwen, DeepSeek distills.Llama-3.1-8B primary, Qwen2.5-7B cross-model, R1-Distill exploratory.
See the benchmarkRead the methods
KODON Weather live - one sensor beating global models, measured skill numbers
Program 03 · Weather · Live

One sensor. Beats the global models where you actually stand.Hyperlocal forecasting with a calibrated uncertainty gate.

The world's forecast systems predict a 6-13 km grid box. We predict your exact spot, and on the ground we beat operational ICON, GFS and ECMWF on the variables that pay. From one station, on a Raspberry Pi, on nine years of real data.Station-level post-processing that outperforms operational ICON, GFS and ECMWF on held-out 2026 data; every number measured, nothing simulated. Same verify-or-abstain gate as the rest of the lab.

3/3global models beaten
9 yrsreal sensor data
1 Piruns the whole thing
  • Tells you when it isn't sure, so a confident forecast actually means something.Explicit uncertainty gate: predict where the signal supports it, abstain where it doesn't.
  • Every number is measured on data the model has never seen.Held-out 2026 evaluation; no in-sample claims.
Open WeatherSee what it beats
Cyber live - free website scan with chained-findings analysis
Program 04 · Cyber Security · Live

Your security team is guessing. We chain the findings.Chained-findings security testing, lab-grade reporting.

Most audits run a scanner, paste the output into a PDF, and charge five figures. Nobody tests what happens when those findings are chained together. We do, with the same rule as the rest of the lab: reproducible, or it doesn't ship.Web, API and LLM attack-surface testing that starts from scanner output and chains findings into real attack paths. Exact-or-abstain reporting: every finding reproducible, severity backed by a working chain.

Freesite scan, live now
Chainednot copy-pasted
0unreproducible findings
  • Start free: point it at your domain and get a real report, not a sales call.Self-serve scan funnel at cyber.koscak.ai; findings triaged before you ever talk to us.
  • Built by the same two people who measure everything else on this page.Same verify-or-abstain methodology as the training and reasoning programs.
Scan my site · freeSee the services
Space project live - NOESIS run on real astronomical data with verified counts
Program 05 · Space & Cosmos · Live

Debugging the universe.NOESIS-AUDIT x DESI: calibrated BAO consistency.

The same engine that finds the law behind a handful of numbers, pointed at the largest dataset there is: the sky. No assumed answers, no cherry-picked signals, structure has to survive held-out validation before we call it structure.A calibrated consistency audit for DESI-class BAO data releases: blind structure recovery over large-scale-structure data, held-out validated, abstaining where the signal does not support a claim.

DESIBAO data releases
blindno assumed answer
held-outvalidated claims only
  • The discovery engine treats cosmological data exactly like any other stream: recover the rule, test it on data it never saw.Same sparse-recovery method as the NOESIS core, applied to BAO consistency checks.
  • When the sky doesn't commit, neither do we, abstention applies at every scale.Calibrated uncertainty on every audit output; no significance-mining.
See the sky projectSee the DESI audit

The team.Authors and operators.

The original scientist, the builder, and the communications layer needed to turn a result into a public research programme.Stochastic neural-network lineage, benchmark engineering, hardware validation, and publication discipline in one compact team.

JK
Dr. Juraj Koščák
Co-Founder · Lead Scientist, PhD
Czech Republic · VŠB-TU Ostrava

PhD (Red Diploma, top distinction) in Computer Science from the Technical University of Košice, Department of Cybernetics & AI, under Prof. Peter Sinčák (founder of the Slovak Artificial Intelligence Society). Doctoral work 2010–2015 pioneered stochastic weight-update methods in neural networks: IEEE WCCI 2010, SCIS&ISIS 2014 (Japan), and the monograph Stochastic Weight Update in Neural Networks (ISBN 3659231029). KSS-LoRA is the direct descendant: the same stochastic-masking principle, transplanted into modern LLM fine-tuning and extended with the Koščák Gamma Theorem, an original result for FP8/FP4 numerical stability.

✉ [email protected] LinkedIn 📄 DBLP ORCID
Filip Koščák
Filip Koščák
Co-Founder · Builder & Research Architect
Europe

Builder and research architect, and Juraj's nephew. Filip runs the lab's engineering end to end: the compute path on real A100 and H200 hardware, the benchmark pipeline behind the 33× result, and the systems that ship. He has the rare ability to see the signal before the data confirms it, and the discipline to prove it before claiming it. A music producer (known on stage as Phauler) turned AI engineer, he cares as much about the craft of the work as the result.

✉ [email protected] 𝕏 @PhilPhauler 📸 Instagram GitHub
LI
Laura Ilcin
PR & Brand Lead
Europe

Laura shapes how KSS-LoRA is seen, and remembered. Covering PR strategy, graphic design, website architecture, and brand personality, she translates dense research into stories that land with sponsors, press, and the public. Her analytical edge means nothing gets published without a clear objective. The reason koscak.ai looks this good.

✉ [email protected]
Back the research

A two-person lab, out-measuring the noise.Independent research needs independent funding.

No big lab behind us, no ad model, no strings. Backing goes straight into compute and hardware so the next result gets measured properly, in the open.Every euro funds GPU time and research hardware. Runs stay reproducible; findings stay public.

€0 raised of €5,00060 days · be the first to back it
€5+
Seed
Fund compute. Get the research updates as they land.
€500+
Hardware
Fund physical research hardware the lab runs on.
€2,000+
Patron
Your name in the paper acknowledgements.
Back this researchView the ALFA round See all tiers

Questions.Common questions.

What is koscak.ai?What is the lab's thesis?+
A two-person research lab building AI that proves instead of guessing. Every program on this page, training, NOESIS, weather, follows one rule: measure it on real hardware, or don't claim it.An independent lab built around verify-or-abstain: systems should quantify what they cannot know and refuse to answer there. One methodology across all programs: controlled runs, held-out verification, exact-or-abstain reporting.
What is NOESIS?What does NOESIS actually do?+
An engine that takes raw data and recovers the rule that generated it, then checks that rule on data it has never seen. When the data isn't enough, it says "I don't know" instead of making something up. You can talk to it live at noesis.koscak.ai.Blind structure recovery: power laws, ODE systems and chaotic dynamics via sparse regression, held-out verified, with a hard abstain boundary. Zero bluffed answers by construction. Live at noesis.koscak.ai.
Is the 33x number real?How is the 33.3x gap reduction validated?+
Yes, and you can check it. 88+ logged training runs on real A100 and H200 GPUs, five seeds, nothing cherry-picked. The full benchmark story lives on the Model Training page.Gap 0.5329 to 0.0160 across 88+ runs, identical configs, 5 seeds, p<0.0001, validated on A100 BF16 and H200 FP8. Full methodology at /training.
Who is behind the lab?What is the research lineage?+
Filip and Dr. Juraj Koščák, nephew and uncle. Juraj spent 15 years turning stochastic learning into published science; Filip builds the systems and runs them on real hardware until the numbers hold.Dr. Juraj Koščák: PhD (TUKE, Sinčák lineage), 2010-2015 stochastic weight-update research, IEEE WCCI 2010, SCIS&ISIS 2014. Filip Koščák: builder-engineer, end-to-end compute path on A100/H200.
Can I try the tools?Which surfaces are publicly live?+
Yes. NOESIS answers live at noesis.koscak.ai, the weather engine runs at weather.koscak.ai, and every research write-up is free on the blog. No accounts, no paywalls.Live: noesis.koscak.ai (query the engine), weather.koscak.ai (calibrated forecasts), /blog (methods + run evidence), /training (full benchmark). Cyber is live at cyber.koscak.ai (free site scan) and the Space program runs live as the NOESIS universe audit.
How is the lab funded?What does backing actually pay for?+
Independently. No big lab, no ad model. The ALFA round funds GPU time and research hardware directly, and every result stays public. You can back it from €5.Every euro goes to compute and hardware; runs stay reproducible and findings stay public. Tiers from €5 (updates) to €2,000+ (acknowledgements). Details at /fund.
Do you collaborate?Is access available for replication or co-authorship?+
Yes, that's the point of publishing. Bring a dataset, a GPU cluster, or a hard question, the two people who built this answer every serious message.Open to replication, compute collaboration, co-authorship, and hardware validation. Direct line to the authors via the contact section below.

Stay with the research.Follow validation updates.

New results, benchmark notes, and publication updates from every program. No noise.Run logs, methodology releases, hardware validation notes, and preprint updates when they are ready.

Open to collaborators, funders, and anyone who checks the mathOpen to replication, compute collaboration, and co-authorship

If you build, fund, or validate AI, let's talk.Replication and compute collaboration welcome.

We answer every serious message ourselves. No funnel, no bots, just the two people who built this.Direct line to the authors. Bring a dataset, a GPU cluster, or a hard question.

Get in touchEmail the lab DM Filip on XReach out on X

References & HardwareSources, targets, and validation surfaces

The hardware and model families behind the lab's results.A100 and H200 are validated; Blackwell-class FP4 is the next validation surface.

Hardware
NVIDIA H200 SXM 141GB

Primary cross-validation GPU. FP8 native. 4.8 TB/s HBM3e. 2.7× faster than A100 for KSS-LoRA workloads.

Hardware
NVIDIA B300 Blackwell Ultra

Next validation target. 288GB HBM3e · 8 TB/s · FP4-class path. Public claims wait for reproducible logs.

Hardware
NVIDIA GB200 NVL72

72-GPU NVLink rack. FP4 native. Koščák Gamma Theorem proves γ_min=1.0, KSS-LoRA satisfies this by design.

Models
Meta Llama 3.1-8B

Primary fine-tuning target. TruthfulQA benchmark. 12 baseline + 40 KSS runs across A100 and H200.

Models
Qwen 2.5-7B

Cross-model validation target. Confirms hardware-agnostic generalization of KSS-LoRA across model families.

Precision
NVIDIA FP8 Training

The precision format that exposes standard LoRA's gradient underflow. KSS-LoRA's result: 5.2% vs 68% quality loss.