Alphius Labs

Quantitative AI HFT Research Platform

A 100% secure infrastructure for AI Agents to test thousands of trading-related ideas per day. Automating the entire pipeline—from hypothesis generation to backtest.

Machine Learning
High Frequency Trading
Low Latency Focus
Rabbit Trader

“when the love of taking risks turns into a lifestyle”

Algorithms

Below is a list of the core algorithms and methodologies powering the platform.

Adapted Monte Carlo Tree Search (A-MCTS)

Redesigned MCTS for optimal parameters search

Imagine you want to explore combinations of over 1000 assets, N targets, M date-ranges, close to 'infinity' number of hypotheses, and close to 'infinity' number of model settings, how would you approach it? Grid search? Random Search? Bayesian algorithms adaptation, Optuna?

MCTS is a complex algorithm that allows to make decisions in difficult situations where full search is impossible. The algorithm is optimized to discover the most perspective scenarios of HFT experiments without human intervention by incorporating HFT specific reward system, compressed memory of previous experiments, and AI agents.

Monte Carlo Tree Search

Statistical Lab

Compare ideas easily

Statistical Lab architecture
Statistical Lab
Statistical Lab console

Statistical Lab is a technology that lets you perform AB tests of your ideas without spending time on experiment design. These are simple examples of questions you can easily answer in the Statistical Lab:

  • Will I maintain the experiment quality and lower the costs of those experiments by using LLM 'Y' instead of 'X'?
  • If I increase the context of a hypothesis generation agent, will it increase the number of SOTA experiments found?
  • Will changing LLM temperature from 0.2 to 0.6 increase my ROI?

Data Lab

The data processing engine is built for HFT scale. For example, Hyperliquid alone generates over 100 GB of data daily. In order to process such massive amounts of data during feature generation and training without RAM exhaustion, lazy mechanisms were developed.

Data Lab Architecture

Completely Safe Insfrastructure

AI Agents operate in strict isolation. They perform the full research cycle without ever accessing the code of trading algorithms, backtest engines, or sensitive data codebases and datasets. No AI model will ever be trained on your intellectual property.

Completely Safe Insfrastructure

Token Optimization

Tokens consumption is one of the most acute problems in agentic systems. The development of compression/search algorithms for coding agent and building no-agentic dependable algorithms to maximize the search space of hyperparameters while minimizing agentic triggers helped to prevent unnecessary token burn during heavy optimization phases.

100k90k80k70k
85k
-15% VS CLAUDE
Alphius Core
100k
Claude Code
94k
Codex
Number of consumed tokens
Scale [70k-100k]

Hybrid Capability

Total flexibility. Agents can either develop custom algorithmic code from scratch, optimize exisitng algorithms, or use the existing high-performance research infrastructure.

Hybrid Capability

Statistical Precision

Advanced mathematical frameworks guide every decision. Manager Agent makes decisions about future actions in conjunction with Bayesian optimization and Thompson sampling to navigate the explore-exploit tradeoff.

Action Selection (Thompson Sampling)

Live Simulation
Probability Density
0.0 (Expected Reward)
1.0
SELECTED
Feature Action
High Uncertainty
Training Action
Medium Uncertainty
Backtest Action
Selected (Max Reward)

Real-time Monitoring

Watch your research ecosystem in action. Different terminals let you track metrics, experiment insights, hardware utilization, and more — all in real time.

Real-time Monitoring

Infrastructure-First AI Research

Alphius Labs builds the critical foundation first, then scales agent freedom responsibly to deliver SOTA results.

Earning Autonomy

The quality of AI research is defined by infrastructure. Giving agents full freedom too early does not usually break code - modern models can often handle implementation. The real failures happen deeper: flawed hypothesis structure, incorrect data processing, memory limits ignored, poor deployment pipelines for HFT production, and model choices that do not match dataset reality.

The agents' autonomy must be earned through engineering solutions and discipline. It is important to solve the most critical infrastructure risks first, because that is what guarantees correctness, repeatability, and trust in outcomes. Once the foundation is reliable, it is possible to expand agent freedom in controlled stages - this is a bottom-up approach in action.

"The goal is to find the point where control and freedom are perfectly balanced. That equilibrium produces the best possible outcomes in any field of research, not just HFT."

Performance vs. Freedom

Research Quality
OPTIMUM
Over-ControlledAgent Freedom →Unbounded

Technologies

vLLM
Ollama
Gemini
ChatGPT
Claude
Jaeger
Prometheus
OpenTelemetry
Alphius Labs Culture

“future is discovered, not predicted”

Contact