Autonomous agents for data analytics
Abstract — We present a scalable, multi-agent architecture for the autonomous orchestration of feature synthesis and machine learning experimentation. By framing the ML lifecycle as a competitive evolutionary simulation, our framework systematically resolves heterogeneous enterprise data into high-fidelity predictive intelligence.
Evolutionary ML Framework
Our proprietary optimization engine treats AI agents as competing populations. Through a continuous cycle of Execution, Evaluation, Selection, and Reproduction, our agents evolve winning model strategies and features without human intervention.
Execution
Parallel agent populations explore distinct segments of the feature space, generating novel hypotheses and code mutations simultaneously.
Evaluation
Automated multi-perspective judges evaluate performance. We assess for innovation, reliability, and edge-case resilience.
Selection & Mutate
Top-performing agents are selected and mutated. Their winning traits are combined to generate the next generation of predictive models.
Use Case: Sports Analytics
Professional sports data provides a highly noisy, unstructured environment ideal for testing automated machine learning. We used this domain to validate our architecture's ability to operate without human intervention.
Our agents autonomously resolved conflicting player identities across different APIs, built concurrent ETL pipelines to ingest raw play-by-play data, and engineered temporal features (like rolling averages and fatigue metrics) at scale. The resulting models successfully predicted complex outcomes like player performance ceilings and injury risk profiles, proving the system's ability to handle raw, messy data end-to-end.
End-to-end autonomous data analytics
Agentic AI replaces the manual hypothesis-experiment-analysis cycle. Our agents discover features, design experiments, and evaluate models across multiple prediction targets.
Autonomous Feature Engineering
AI agents generate predictive features from multi-source enterprise data autonomously. A novel dual lag schema captures both weekly and per-game temporal patterns, scaling feature spaces far beyond what manual engineering achieves.
Automated ML Experimentation
End-to-end pipeline that generates hypotheses, designs experiments, tunes hyperparameters via Bayesian optimization, and evaluates results across multiple prediction targets and player positions — no human in the loop.
Autonomous Semantic Unification
A multi-layered pipeline ingests, validates, and aligns heterogeneous data sources utilizing zero-shot LLM entity resolution, achieving high-fidelity cross-source mappings without rigid deterministic rules.
From raw data to predictions
A layered pipeline where each stage transforms, validates, and enriches — fully automated, end to end.
Unstructured Ingestion
Raw heterogeneous enterprise data is ingested and subjected to autonomous schema inference.
Semantic Resolution
Cross-domain entities are unified via zero-shot probabilistic mapping, bypassing legacy ETL logic.
Autonomous Feature Synthesis
Agent swarms traverse high-dimensional spaces to synthesize and validate non-linear temporal features.
Agentic Experimentation
Decentralized execution of hypothesis generation, Bayesian tuning, and multi-metric model evaluation.
Predictive Biometrics & Markets
Longitudinal inference modeling biomechanical fatigue, performance ceilings, and dynamic risk profiles.
Active research areas
We work on genuinely unsolved problems at the intersection of autonomous AI and complex data environments. Overcoming these technological bottlenecks requires systematic investigation, where failed hypotheses are a necessary byproduct of pushing the boundaries of applied machine learning.
- Uncertainty 01Autonomous Feature Synthesis in High-Dimensional SpacesCan multi-agent systems autonomously generate, test, and validate non-linear feature interactions without relying on human domain-knowledge priors? We are exploring the mathematical ceiling of automated feature discovery and cross-season generalization without overfitting.
- Uncertainty 02Agentic Orchestration of the ML LifecycleCan a decentralized swarm of AI agents reliably replace the human hypothesis-experiment-evaluation loop? We are engineering frameworks where autonomous agents manage computational budgets, avoid local optima, and orchestrate complex ML pipelines end-to-end.
- Uncertainty 03Semantic Entity Resolution via LLMsCan heterogeneous, unstructured temporal data be automatically unified with ML-grade fidelity? Standard deterministic logic fails across diverse complex data sources. We are researching autonomous schema alignment and zero-shot entity resolution to solve this at scale.
- Uncertainty 04Evolutionary AI Code MutagenesisCan AI systems safely self-optimize their own underlying architecture? We are building an evolutionary framework where AI "civilizations" mutate codebases, compete, and improve through automated, multi-perspective judged evaluation without human intervention.
- Uncertainty 05Non-Standard Target Prediction via Multi-Task LearningCan deep learning models accurately map highly complex, non-differentiable human targets? We are utilizing quantile regression and temporal graphs to predict longitudinal outcomes—such as biomechanical fatigue thresholds, performance ceilings, and recovery optimization.
- Uncertainty 06Compute-Constrained Agentic SwarmsAs autonomous agents scale, can inference costs be dynamically optimized without degrading reasoning quality? We are actively researching dynamic routing—directing routine pipeline tasks to specialized local models while reserving heavy reasoning models strictly for complex architectural decisions.
Engineering Team
Complementary expertise in multi-agent architectures, semantic resolution, and evolutionary computation.
Olivier
Architect of the autonomous semantic resolution systems and agentic ML experiment frameworks. Designed the feature engineering schema and pioneered the agentic AI methodology for autonomous experimentation.
Garry
Built the AI development infrastructure — agent workflows, swarm delegation, ML auditing, observability. Designed the evolutionary AI framework where agents compete, mutate, and evolve code.
Researchers, engineers, or organizations interested in our autonomous architecture and evolutionary ML frameworks. Our principal investigators are open to technical discussions.