3. Data, Science & AI

Integrates Claude AI agents for multi-agent training and live trading systems, including orchestrator configuration and safety mechanisms.

Provides GPU-safe parallel processing patterns for KINTSUGI to prevent OOM crashes and deliver Jupyter-compatible progress updates during data-intensive computations.

Determines optimal neural network architecture for PPO trading models by balancing capacity, inference speed, and hidden layer configuration.

Provides a 7-action position sizing space (25/50/75%) for small account simulation in trading models, triggered during model training and observation dimension upgrades.

Efficiently caches historical market data using gap-filling to prevent redundant API calls and enable incremental updates.

Verifies AI-generated claims against uploaded documents using DeepCitation for contradiction detection and validation.

Trains reinforcement learning models by resampling data across multiple timeframes, such as 1-hour and 4-hour intervals.

Fixes PPO training early-stop issues by handling impossible drawdowns (>100%), preventing premature termination, and enabling adaptive recovery.

Fixes KeyError: 'uid' in Plotly FigureWidget during interactive dashboard shape updates with sliders.

Integrates spatial proteomics (CODEX) and single-cell RNA sequencing (scRNA-seq) data through region-aware cell matching for multi-omics analysis.

Automates explicit GPU memory cleanup using CuPy in Jupyter to prevent cell hangs during data processing and AI model execution.

Excludes dotted symbols (e.g., warrants, units, class shares) from yfinance data to prevent errors in sector lookups and training notebook execution.

Calibrates reward_scale to fix phantom MaxDD values and align MaxDD with training quality in reinforcement learning for trading strategies.

Tunes parameters for ecological predator-prey simulation models, including regime detection and cooldown mechanisms.

Provides historical financial data access and analysis for algorithmic trading strategies using Alpaca Algo Trader Plus and yfinance comparisons.

Filters candidates by sector using yfinance data to manage pool size, ensure diversity, and reduce historical data download time.

Evaluates caching for BaSiC illumination correction in microscopy, noting limitations for sparse marker datasets.

Accelerates correlation matrix calculations using GPU and SQLite caching to eliminate bottlenecks in symbol selection.

Manages model training archives and implements gating (APPROVED/REVIEW/DROP) for deployment decisions, historical reference, and overfitting mitigation.

Aligns RNA and protein data modalities via z-score normalization for cross-modal integration, requiring both to be z-scored.

Dynamically adjusts MaxFuse parameters for protein panel size variations (26 to 59+ markers), optimizing data processing performance.

Prioritizes GPU processing to maintain quality without speed trade-offs in AI/ML workloads, ensuring optimal performance.

Automates saving quantification quality control heatmaps and profiles as PDF files for KINTSUGI data analysis workflows.

Optimizes segmentation, SOM clustering, and spatial analysis for multiplex immunofluorescence imaging with cell compartment and ECM feature extraction.