Skip to main content
Don’t miss what’s next at Microsoft Ignite
Don’t miss what’s next at Microsoft IgniteGet direct access to the people, products, and expertise behind Microsoft technology, with opportunities to ask questions, build skills, and learn alongside the community.
Register now
Microsoft
separator
https://catalogartifact.azureedge.net/publicartifacts/fractal.pievolve-09a0dd2a-7201-41c5-80cf-2970966d09dd/image7_logo216216x216.png

PiEvolve - Intelligence that Evolves

by Fractal Analytics Inc.

PiEvolve - Intelligence that Evolves

1. Problem statement

Enterprise data science teams spend 60–70% of their effort on the least differentiated part of their work — EDA, cleaning, preprocessing, feature engineering and iterative tuning — while the models they were hired to build wait. A firm with 150 data scientists at an average $151K total comp burns roughly $17M a year on preparation alone. But the deeper problem is not throughput, it is context loss. Real business problems are solved by domain and business context: what the target should actually be, which segments matter, which metric reflects commercial reality, what counts as leakage versus legitimate signal. In traditional working models that context lives in people's heads and gets transferred through hand-holding — briefing calls, review meetings, tribal knowledge — and in discrete tool-based workflows it fragments further, because every hand-off between data prep, feature engineering, modelling and reporting drops the reasoning behind the previous step. What survives is an artifact; what is lost is why. One-shot pipelines and generic agents make this worse: they answer the ask but miss the drivers, they analyze data rather than reason over it, and they are not tuned for real-world, large-scale data science work. The result is technically valid models that are commercially uninteresting, and insight that never becomes decision-ready.

2. How the solution addresses the problem

PiEvolve fixes both halves — it keeps the context and it keeps searching. Its Domain Context Protocol (DCP) is not a plugin but a standard for embedding enterprise domain knowledge, letting the system extract information hierarchically and in depth, and grounding every data operation, problem framing and metric choice in the client's domain — reusable, across use cases, teams and projects. Because preparation, evolution and reporting all run inside one persistent framework, nothing is handed off: the reasoning that shaped a feature is still available when the model is selected, and still available when the business report is written. On top of that grounding sits evolutionary optimization. Rather than one-shot model building, PiEvolve performs graph-structured optimization over solution trajectories, abstracting business logic into executable NodeChains that keep evolving until the compute budget is exhausted, with priority-based sampling with decay to avoid local optima and a dual strategy that refines winning branches while aggressively debugging failures to lift the performance floor. The outcomes track the claim: a 26-week pharma forecast across 16 series reached SMAPE 4.26 against a 26.78 naive baseline — an 84% improvement — in ~2 hours, 51 autonomous reasoning steps, $13.23 of inference and zero data scientist time. A used-vehicle Days-to-Sell problem (505K rows, 14.6K VINs, 66 features) yielded a 12–20 day DTS reduction worth $2.5M–$7.0M in annual carrying-cost savings, with drivers split into what dealers can act on versus what they should monitor. Scaled across Fractal's ~700 data scientists, the same efficiency implies ~$22M a year.

3. Product description

PiEvolve is a persistent evolutionary agentic framework for autonomous machine learning, combinatorial optimization and scientific discovery — intelligence that evolves, grounded in your domain. It manages the end-to-end lifecycle in one place. A Data Preparation Agent turns messy enterprise data from databases, warehouses and files into ML-ready data at real-world scale; a Research Agent performs grounded deep research over the problem; an ML Evolution Agent explores and evolves the solution space through long-horizon optimization, with PiNotebook for live code interaction mid-run; and an Insights & Reporting Agent converts the accumulated run context — reasoning traces, feature importance, search logs, validation folds — into both an ML Insights Report for engineers and a Business Action Report for leaders. Two pillars make it work: the Domain Context Protocol for embedding enterprise domain knowledge, and evolutionary optimization across ML and combinatorial problem classes, supported by data-science tooling built for large-scale data that keeps agent cost down. It runs autonomously or collaboratively, supports pause/resume for long jobs, integrates into existing pipelines, and is benchmarked on Open AI's MLE-Bench. It is delivered three ways: as a standalone license deployed in the client environment, via Forward Deployed Data Scientists certified on the platform, or bundled into outcome-based solution delivery.

At a glance

https://catalogartifact.azureedge.net/publicartifacts/fractal.pievolve-09a0dd2a-7201-41c5-80cf-2970966d09dd/trailer_4021755009991186890_trailer.png
/staticstorage/20261108.1/assets/videoOverlay_62a424ca921ff733.png
https://catalogartifact.azureedge.net/publicartifacts/fractal.pievolve-09a0dd2a-7201-41c5-80cf-2970966d09dd/image5_SS1pievolvechatagentview.png
https://catalogartifact.azureedge.net/publicartifacts/fractal.pievolve-09a0dd2a-7201-41c5-80cf-2970966d09dd/image3_SS2pievolvenodeexpansionview.png
https://catalogartifact.azureedge.net/publicartifacts/fractal.pievolve-09a0dd2a-7201-41c5-80cf-2970966d09dd/image1_SS3pievolveoptimisationdashboard.png
https://catalogartifact.azureedge.net/publicartifacts/fractal.pievolve-09a0dd2a-7201-41c5-80cf-2970966d09dd/image8_SS4pievolveoptimisationtree.png
https://catalogartifact.azureedge.net/publicartifacts/fractal.pievolve-09a0dd2a-7201-41c5-80cf-2970966d09dd/image4_SS5pievolvesolutionstrategygraphs.png
English (United States)
Your Privacy Choices Opt-Out Icon Your Privacy Choices
Consumer Health Privacy Sitemap Contact Us Privacy & Cookies Terms of Use About our ads Manage cookies