Top Builders

Explore the top contributors showcasing the highest number of app submissions within our community.

Kraken

Kraken, operated primarily under the corporate umbrella of Payward Inc., represents a highly evolved shared financial infrastructure layer and digital asset exchange. Founded in 2011, the platform provides unified liquidity, centralized risk management, and a globally compliant infrastructure to serve diverse client segments.

General
Provider / AuthorKraken
Founded2011
TypeHigh-Performance Digital Asset Exchange & Financial Infrastructure
Primary ArchitecturePolyglot Microservices (Rust gRPC backend, C++ matching engine)
Websitehttps://www.kraken.com/en-gb/lp/platform
Developer Documentationhttps://docs.kraken.com/
Open-Source Repositoryhttps://github.com/krakenfx

Start building with Kraken products

Kraken provides an institutional-grade trading ecosystem capable of sub-2 millisecond round-trip latencies and handling up to 15,000 algorithmic orders per second. The underlying architecture utilizes a highly optimized microservices topology, with the ultra-hot path matching engine engineered in C++ for deterministic execution and the backend primarily built in Rust. Furthermore, Kraken offers developers the "Ink" Layer-2 blockchain, a scaling solution built on the Optimism OP Stack that seamlessly bridges centralized exchange liquidity with decentralized finance (DeFi). We highly encourage you to check out the innovative Web3 apps and algorithmic trading bots created with this technology during lablab.ai hackathons!

Key Platform Features

  • Polyglot Microservices: Utilizes Rust and the Tokio asynchronous runtime for highly efficient network I/O and routing, while relying on Python (Django) with strict static type checking (mypy) for API layer logic.
  • Deterministic Latency & Colocation: Offers dedicated colocation services through partnerships with the Beeks Group, allowing institutional traders to physically host their servers in the same data centers as the matching engine.
  • Dynamic Rate Limiting: Employs a sophisticated mathematical decay model that assigns penalty points based on computational burden and order resting time, rewarding efficient algorithmic behavior and detering malicious spoofing.
  • Cryptographic Proof of Reserves: Uses a mathematically verifiable Merkle Tree data structure, independently attested by a CPA firm, enabling clients to cryptographically verify their included balances without compromising privacy.
  • The Ink Layer-2 Blockchain: An optimistic rollup anchored to Ethereum that provides deep integration with Kraken's centralized infrastructure, empowering frictionless transitions to decentralized applications (dApps) without relinquishing private key custody.

Libraries and SDKs

  • API Specifications (api-specs): Machine-readable open-source specifications utilizing OpenAPI for REST and AsyncAPI for WebSocket channels.
  • Golang Implementation (api-go): An officially maintained, highly concurrent Go module designed for rapid interaction with Spot and Derivatives API endpoints.
  • Python WebSocket Client (kraken-wsclient-py): A native Python client optimized for establishing robust, persistent WebSocket connections for real-time order book feeds.
  • python-kraken-sdk: A community-driven Python SDK featuring built-in handlers for automated nonce calculation, REST retries, and asynchronous handling of WebSocket feeds.
  • KrakenTools: A public utility toolkit offering mathematical helper functions for cryptographic payload signing and data structuring.

Boilerplates

  • Decentralized Web Wallet (wallet): A self-custodial wallet framework written entirely in TypeScript, foundational for Web3 integration.
  • Futures Trading Bot Template: A deeply documented template demonstrating complex algorithmic execution and advanced error handling.
  • Freqtrade Integration: An open-source crypto trading bot allowing developers to test and deploy machine learning models and statistical arbitrage strategies live against the Kraken API.
Edit on GitHub

kraken AI Technologies Hackathon projects

Discover innovative solutions crafted with kraken AI Technologies, developed by our community members during our engaging hackathons.

AI-Driven 3-Phase Digital Twin

AI-Driven 3-Phase Digital Twin

Industrial operations heavily rely on the continuous performance of 3-phase induction motors. Unexpected failures caused by phase unbalance, stator overheating, or mechanical degradation can lead to catastrophic downtime and expensive repairs. To address this, our project introduces the "Intelligent SCADA Digital Twin," a real-time telemetry monitoring and predictive maintenance system. The architecture features a backend engine generating highly realistic 3-phase electrical loads (Current R, S, T), voltage sags, power factor variations, and mechanical vibration metrics. At the core of our solution is an AI-driven predictive maintenance model utilizing the Isolation Forest algorithm. It continuously analyzes the streaming telemetry to identify critical anomalies. When severe imbalances or thermal thresholds are breached, the AI instantly triggers a simulated PLC (Programmable Logic Controller) interlock signal to safely trip the circuit before physical damage occurs. The backend is powered by FastAPI for high-speed data delivery, while the frontend utilizes Streamlit to render a professional industrial HMI panel. Although deployed on standard compute nodes due to global platform demand, the data processing pipeline is architecture-agnostic and fully optimized to leverage the computational power of AMD EPYC processors and MI300X accelerators. Moving forward, this prototype's framework is planned to be adapted for the predictive maintenance of railway Point Machines (Motor Wesel), directly enhancing national transportation safety infrastructure.

ChiefFlow AI — AI Chief of Staff

ChiefFlow AI — AI Chief of Staff

ChiefFlow AI is an AI Chief of Staff that automates the operational grind every small business faces: reading invoices, reviewing contracts, triaging support tickets, scheduling meetings, and replying to email, all without losing human oversight where it matters most. A Manager Agent classifies every incoming item, whether an email, PDF, pasted text, or API call, by intent: invoice, contract, complaint, tender, meeting request, or support ticket. It then routes the task to one of six specialist agents: Email, Finance, Legal, Research, Calendar, and Support. Each agent extracts structured data and drafts a recommended action. High-risk actions like payments, contracts, and external communication always pause for one-click human approval before executing. Every step is logged to a full audit trail. Rather than sending every task to the biggest model available, ChiefFlow AI routes by task complexity: simple tasks to Gemma, moderate tasks to an AMD GPU-hosted open model, and complex reasoning like contracts and tenders to Fireworks AI, all accessed through Fireworks' AMD-hosted infrastructure. If a tier is unavailable, the system gracefully degrades to a deterministic local reasoning engine, so it is never broken by a missing key or network issue. Built as a full-stack, containerized application: FastAPI backend with real agent orchestration, a Next.js frontend styled as genuine enterprise SaaS, SQLite persistence, live analytics, and a complete audit log. Ships as a single Docker image, one command to run anywhere.

ROCmPorter Agent

ROCmPorter Agent

ROCmPorter Agent is a local-first assistant for teams moving CUDA and NVIDIA-heavy projects toward AMD ROCm. It scans GitHub repositories for CUDA source files, CUDA headers, NVCC/build-system assumptions, NVIDIA container signals, and GPU package risks. The developer can choose an evidence file and generate a single-file ROCm review artifact using a local Ollama coding model, so the workflow does not require paid cloud LLM APIs. The product does not blindly apply AI output: it stores the diff, validates syntax where local tooling exists, replays the patch, checks for source drift, scores review risk, and blocks unsafe workspace apply until verification is strong enough. It exports offline JSON, Markdown, HTML, diff, zip, checksum, verification receipt, and GitHub-ready review artifacts. A tracked submission proof run completed 3/3 pinned CUDA build-system cases with 3 export-ready review artifacts, 0 export blocks, 0 infrastructure failures, and 0 high-risk patches.

TokenOptimizer — Hybrid Routing Agent

TokenOptimizer — Hybrid Routing Agent

TokenOptimizer is a hybrid, token-efficient routing agent for Track 1. It treats the leaderboard as what it is — an accuracy gate followed by a race to the fewest Fireworks tokens — and optimizes directly for that. Every task first hits free, deterministic solvers: an exact arithmetic/percentage/average calculator, a transitive-ordering solver for comparative logic puzzles, and a syllogism checker. Whatever they can prove is answered for zero tokens. Remaining tasks go to a small local model (Qwen2.5-Coder-3B via Ollama, baked into the image) whose answers are checked with free verifiers — valid number, parseable JSON, compilable code, real sentiment label, length constraints — plus self-consistency sampling. Only genuinely uncertain tasks escalate to a Fireworks model, using a ~3-token system prompt, compressed input, a capped max_tokens, and reasoning_effort=low to keep completion tokens minimal. The solvers are red-teamed to never emit a wrong answer — they defer to the model whenever unsure — so the accuracy gate is protected while local-first routing drives token usage toward zero. It ships as a self-testing linux/amd64 Docker image that reads ALLOWED_MODELS from the environment and handles all eight capability categories: factual Q&A, math, sentiment, summarization, NER, code debugging, logic, and code generation.

StyleCap

StyleCap

StyleCap is a containerized video captioning agent built for the AMD ACT II Video Captioning track. Given any video clip, it produces four captions of the same scene in four distinct tones: formal, sarcastic, humorous with tech references, and humorous with everyday humor. Captions stay faithful to what is actually on screen while nailing the requested voice. The pipeline is simple and fast by design. The agent reads tasks from /input/tasks.json, downloads each clip, and samples 8 evenly spaced frames in memory with OpenCV, downscaled to 768px. A single multimodal call to a vision language model (kimi-k2p6 served by Fireworks AI) generates all four styles at once, guided by few-shot tone exemplars promoted through our own LLM-judge evaluation harness. Four parallel workers process clips concurrently, so the three official example clips finish in about 20 seconds end to end.Reliability comes from layered fallbacks: API retries, then per-style recovery calls, then guaranteed generic captions, so every requested style is always present and a broken URL never sinks the run. The output is written to /output/results.json in exact harness format with exit code 0. Nothing is hardcoded to the example clips. Prompts are purely style driven, so the agent generalizes across nature, urban, animals, people, sports, food, weather, and technology footage. The image is linux/amd64, 1.21GB, MIT licensed, and publicly pullable from GHCR.

SplitLLM – Explainable Hybrid AI Routing

SplitLLM – Explainable Hybrid AI Routing

SplitLLM is a production-inspired hybrid AI routing platform designed to make large language model deployments faster, more cost-efficient, and fully explainable. Instead of sending every request to a single cloud model, SplitLLM first analyzes each prompt using a multi-stage reasoning pipeline to determine its complexity, confidence requirements, expected cost, and execution strategy. Based on this analysis, it intelligently routes requests to either lightweight local models or powerful cloud models, ensuring the right model is used for the right task. Unlike traditional routers that only decide where to send a request, SplitLLM explains why every routing decision was made. Every prediction includes transparent reasoning, confidence estimates, routing factors, telemetry, and performance metrics, enabling users and organizations to understand and trust the system's decisions. The platform includes modular routing, configurable budget policies, live telemetry, benchmarking tools, a reasoning profiler, CLI utilities, REST APIs, Docker support, automated testing, and CI/CD workflows. It is designed with production engineering principles, making it easy to deploy, monitor, and extend. Built with AMD's AI ecosystem in mind, SplitLLM supports efficient local inference workflows and hybrid deployment strategies that reduce unnecessary cloud usage while maximizing available compute resources. The architecture is modular, scalable, and optimized for future integration with additional models and routing strategies. Our vision is simple: AI should not always use the biggest model—it should use the smartest one. SplitLLM delivers intelligent, explainable, and efficient AI routing for modern applications while helping developers reduce operational costs, improve response times, and maintain high-quality outputs.