Products & Tools
Sun Research Lab is actively developing research tools and software prototypes that translate our work on bias dynamics, AI fairness, and systems engineering into practical, usable instruments. All tools are currently in active research and development.
BiasFlow is a testing and auditing framework for detecting, tracing, and measuring bias as it propagates through multi-stage machine learning pipelines. Grounded in our state-based theoretical framework for stage-aware bias propagation, BiasFlow aims to give practitioners a systematic way to identify at which pipeline stage bias is introduced, amplified, or becomes undetectable.
The prototype is currently being built alongside our manuscript submissions to UAI, ICML, and ASE (2026). The tool is designed to operate on existing ML pipelines with minimal instrumentation overhead.
Stage-aware tracing
Instruments each pipeline stage independently to localize where bias enters or amplifies.
Uncertainty quantification
Models interactions between bias and uncertainty to identify hidden or compounding effects.
Audit reports
Generates structured reports mapping bias scores to pipeline stages for interpretability.
Pipeline-agnostic
Designed to wrap existing ML workflows without requiring architectural changes.
Hybrid Search
A research tool combining dense vector search with traditional lexical retrieval methods, designed to improve search quality and fairness in AI-powered information systems. Under active study to understand how retrieval bias manifests across hybrid ranking strategies.
DVFS Testing Framework
A framework for evaluating Dynamic Voltage and Frequency Scaling (DVFS) policies in real-time AI task scheduling environments. Builds on published schedulability analysis work to create a testable, reproducible benchmark for power-performance trade-offs in IoT and edge AI.
Fairness Constraint Propagation
A research system for encoding and propagating fairness constraints through machine learning pipeline stages — ensuring that fairness guarantees specified at one stage are preserved, transformed correctly, or flagged when violated at subsequent stages.
BiasFlow Testing
The bias testing methodology underlying the BiasFlow prototype — a principled approach to writing test cases that surface bias amplification and identifiability failures in ML pipelines. Conceptually related to data dependence-based testability transformation from prior software reliability research.
Prototype ActiveTool development stages — 2025–2026
Interested in collaborating on any of these tools? We welcome research partnerships and early access inquiries.
Get in touch →