Product
Inductus
Multi-agent intelligence for quantitative research and data science—reasoning, orchestration, and workflows built for your desk.
Built for desks that test hypotheses, not slogans
Inductus is aimed at quantitative researchers, fund analysts, and data scientists who need more than a generic assistant: a structured environment where natural language turns into traceable steps, data you already own sits next to curated fundamentals, and every session can be exported into notebooks your team can audit and reproduce.
The sections below spell out how the agent orchestrates work, grounds knowledge, and persists context—so you can judge fit before you launch a workspace.
Workspace and workflow
The product centers on a single workspace: conversational control, dataset discovery, and paths to notebook export in one surface. You steer the agent with prompts; it proposes decomposed steps instead of opaque completions, which keeps complex asks—like covariance work or multi-table joins—inspectable as a chain of actions rather than a black box.
Sessions stay transparent: you can follow what ran, what data it touched, and how outputs were produced, then push results into Jupyter-style workflows when you are ready to harden methodology or share with a desk.
Multi-agent orchestration and models
Behind the scenes, prompts are broken into sequential, executable steps—autonomous task chaining—so “one big ask” becomes a legible pipeline you can interrupt, correct, or rerun. That matters when a single slip in ordering or data selection would invalidate a research conclusion.
Model choices emphasize precision and latency discipline: the stack leans on small, task-focused language models (SLMs) where they outperform monolithic “chat” LLMs for structured finance and data work—reducing both cost and the failure modes of opaque general-purpose endpoints.
Knowledge extraction and memory
GitHub-to-logic mapping lets you point the system at repositories so methodologies in code can be lifted into the workspace—not just summarized, but connected to executable context.
IR triplet learning turns raw notebooks into structured, searchable “knowledge triplets,” so prior analysis does not disappear into static PDFs or orphaned cells.
Recursive memory keeps a persistent view of how experiments evolve: the agent is shaped by what you actually ran, not a one-off chat transcript.
Together, these pieces push beyond casual chat: the goal is a scientifically grounded loop—hypothesis, instrumentation, observation, revision—rather than anecdotal answers.
Data and integrations
Quant teams routinely burn most of their time on plumbing—cleaning OHLCV, aligning identifiers, reconciling vendor schemas—before they ever test a signal. Inductus is designed to collapse part of that tax: automatic discovery and cataloging of local CSV and Parquet, with the agent reasoning over structure and joins in conversation.
Where it fits your process, built-in access to Sharadar SEP / SF1 fundamentals supports exploratory work and method validation in a controlled setting—alongside your own files, not instead of them.
The intent is simple: spend less time mapping columns and more time deciding whether an idea survives contact with the data.
Disclaimer: Inductus AI is not a stock picker or trading advisor—it is a tool for exploring data and testing hypotheses.