AI Enterprise SolutionsGPT & Claude

QALA: secure app streaming that stops AI cloning

WHO IT'S FOR: software vendors (ISVs) that sell desktop apps (accounting and ERP, CAD and engineering tools, industry software), and companies with internal apps they must keep private. WHAT GOES WRONG TODAY: every install ships the app's binary to the customer's PC. Open-source LLM decompilers turn binaries back into working code (LLM4Decompile-9B-v2, 2024: 64.9% of functions re-execute correctly), and AI agents can drive any desktop UI and map an app screen by screen. Obfuscation only slows them down, so a product built over years can be cloned in days. 8+ app-streaming platforms exist (AWS, Microsoft, Citrix, Omnissa, Google, Kasm...) but none of them detects AI cloning. OUTCOME WITH QALA: the app runs in QALA's cloud and users receive only pixels, never the code. Our AI Detection Engine scores every session's behaviour live (speed, timing, coverage, forward sweeps, pointer, hidden honeypot screens) and monitors, asks for re-authentication or locks the session when an agent tries to copy the app through its screens. End users need no costly hardware or installs and can work from any device. WHY IT PAYS: for a 500-user ISV with $5M revenue, the Pro plan ($18/user/month, $108k a year) pays for itself if it prevents cloning losses of 2.2% of revenue (0.4% counting support savings).

Open the demo Video GitHub

Preliminary score

71.5/ 100Strongaverage of 2 judges

Value for the user17.0 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI19.0 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing16.5 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility12.0 / 15
Data requirements, running costs and a clear next step.
Originality7.0 / 10
A meaningful difference from common solutions.

What the judges wrote

Claude Opus 5.5Preliminary74/100
Strengths
Your testing is unusually honest and systematic. You compare against a realistic baseline (a rate-limit rule), hold out seed 777 from training, average over 30 seeds, and openly report where v1 fails (3% on low-and-slow agents) and the late-lock problem. The v2 model is genuinely trained by you: it is reproducible via train_v2.js, and 16 tests check that the demo runs the same code as the tested module, which makes your numbers credible. Your feasibility story is concrete, with gateway-metadata-only data needs, $5–6 per user infra cost, a 90-day plan and an explicit pivot rule.
To improve
Every attacker and every 'real user' comes from your own simulator, so the detector may simply be learning your generator's quirks. Even a handful of real human sessions and one real agent (for example a browser-use or computer-use agent) mapping the app would make the results far more convincing. The streaming layer is a Web Worker simulation, so a minimal Guacamole session with the detector attached would show the core enterprise workflow end to end. Be consistent about false positives: v2's 30-seed average of 3.5% is worse than v1's 2.4%, yet the deck leads with 1%. The threat of cloning an app through its UI also needs evidence from ISV interviews, not only the decompiler statistic.
GPT-6 AstraPreliminary69/100
Strengths
Your behavioural detector gives AI a concrete role, with a reproducible training process and clear separation between the streaming preview and planned production integration. Your testing stands out for multi-seed baseline comparisons and specific failures, including low-and-slow evasion and locks occurring after roughly 80% screen coverage. Your pilot roadmap includes a useful customer-validation threshold and an explicit pivot rule.
To improve
Your next prototype should stream an actual desktop app through a server-side gateway and enforce detector decisions there. Validate your detector on real users and adaptive agents, reporting false-positive rates and screen coverage before intervention—not just eventual detection. Narrow your 'stops AI cloning' claim to demonstrated protections, and validate the loss-prevention assumptions behind your ROI with ISVs. Measure concurrent-session compute and bandwidth costs to substantiate your $5–6 per-user monthly infrastructure estimate.

The team

Fuadkhan Safarov
developer
Ali Shaban
developer
Ali AghayevLead
developer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: v1: transparent weighted logistic classifier over 7 behavioural features (speed, timing regularity, screen coverage, forward sweep, pointer movement, backtracking, dwell) plus a honeypot rule, hand-calibrated. v2 (trained by us, train_v2.js in the repo): averages v1 with a logistic-regression model trained on 10,612 labelled session prefixes from 1,000 simulated sessions (seed 4242), threshold checked on seed 9191, seed 777 held out for testing. No hosted or third-party AI model is called at runtime. AI tools: Claude (Anthropic) used as a coding and writing assistant.
Data: Synthetic sessions generated in the browser by our simulator (real-user model plus 4 attacker types), plus the visitor's own clicks in the demo. No personal or external datasets. In production the detector needs only gateway event metadata (screen, action, timestamp, pointer presence), never screen content.
Components: Vanilla JavaScript, Canvas, OffscreenCanvas and a Web Worker (no libraries); Node's built-in test runner. Planned production streaming layer: Apache Guacamole (Apache-2.0). Fonts: Google Fonts (Space Grotesk, Mona Sans, Chivo Mono). Templates: pitch deck built on a Canva template; website style inspired by the CyberGuard Webflow template (Flowzai), rebuilt from scratch with no template code or images. Video: recorded with Playwright; voiceover generated with Piper TTS (open-source, en-US 'ryan' voice); music and sound effects synthesized with NumPy; encoded with ffmpeg. Hosting: Vercel.
Built during the event: Confirmed

Quality testing

SETUP: 100 simulated sessions per attacker type, seed 777 (held out from training), compared with today's common defence, a rate-limit rule (more than 25 actions per minute). CAUGHT (rate limit / QALA v1 / QALA v2): - Naive crawler: 100% / 100% / 100% - Jittered bot: 100% / 100% / 100% - Human-mimic agent: 43% / 67% / 73% - Low-and-slow agent: 1% / 3% / 38% REAL USERS WRONGLY FLAGGED: 18% / 1% / 1% AVERAGED OVER 30 RANDOM SEEDS (v1 / v2 / rate limit): human-mimic caught 63% / 69% / 40%; low-and-slow caught 4% / 29% / 0%; real users flagged 2.4% / 3.5% / 17%. AUTOMATED TESTS: 16 pass (node --test): determinism, the published v1 and v2 numbers, response thresholds (45 / 70 / 85), honeypot locking, a calm human session staying Normal, the multi-seed comparisons, reproducing the trained weights, and a check that the demo page runs exactly the same code as the tested module. WHAT BROKE: 1) Low-and-slow agents beat v1 (3% caught). The trained v2 raises this to 38%, but it leans on the pointer signal, which a real agent can fake. Next: cross-session limits and pilot data. 2) Late locks: a caught mimic has often mapped about 80% of the screens before the lock. Next: earlier step-up and more honeypots. 3) About 1% of fast real users get a step-up prompt. Thresholds are tunable per app. LIMITATION: all sessions are simulated. Our next step is pilots with local ISVs to measure the detector on real users and real agents.

Worth 20 of the 100 points.

Running it

No setup and no login: open the demo link in Chrome, Edge, Firefox or Safari on any laptop. Website: https://qala-ai-detector.vercel.app/ 1) Live console: click around the sample business app and watch the risk score and the response level (Normal / Monitor / Step-up / Lock). 2) Streaming preview: the app is rendered by a simulated host in a Web Worker and arrives as watermarked JPEG pixel frames; your clicks go back as coordinates only. 3) Attack lab: launch a naive crawler, a jittered bot, a human-mimic agent or a low-and-slow agent and watch the detector respond. 4) Benchmark: rerun 100 sessions per attacker type with any seed and compare v1, the trained v2 and a rate-limit rule. Tests: clone the repo and run `npm test` (Node 18+, no dependencies). `npm run train` reproduces the trained v2 weights.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.