Skip to content
Research Assistant

Research that shows its sources.

Ask a question, upload documents, or describe your ideal client. You get a structured report where every claim is marked as verified, estimated, or inferred, with the source and date beside it.

General, technical, business, and document research, plus B2B lead discovery, in one workspace.

Compare LangGraph and CrewAI for production agents

  1. LangGraph models a workflow as a stateful graph, with conditional routing between steps. [1]

  2. Graph state can be checkpointed, so a paused run resumes where it stopped. [3]

  3. For workflows that need manual review steps, the graph model maps more directly. [1, 3]

  4. Migrating a small linear chain is likely days of work, not weeks. [no source]

[1] LangGraph docs, retrieved 26 Oct[3] Persistence guide, retrieved 26 Oct
Example output. Each mark shows how far a claim can be trusted.

Every claim tells you how much to trust it.

Reports, lead profiles, and document answers all use the same five marks, so you can tell fact from guesswork without reading the fine print.

Verified fact
Stated by a source you can open. The passage that backs it is highlighted.
Estimate
A reasoned figure or range. Labelled so it is never read as a fact.
AI inference
A conclusion drawn from the evidence. The sources it draws on are listed.
Needs verification
Found, but weakly sourced or out of date. Check it before you act on it.
Not found
Searched for and not found. Left blank, never filled with a guess.

One way of working for every kind of question.

Each research type gets its own source priorities. All of them save to projects, so you can continue a line of research next week and compare findings across sessions.

  • General research

    Breaks a question into smaller searches, compares sources, and flags where they disagree.

    “What changed in Australian privacy law for small businesses in the last two years?”

  • Technical research

    Ranks official documentation, specifications, and source repositories first.

    “LangGraph vs. CrewAI for a production agent with manual review steps.”

  • Business research

    Markets, competitors, and business models, with every company claim dated and sourced.

    “Opportunities for an AI automation consultancy targeting Australian SMBs.”

  • Document research

    Upload PDFs, text, and Markdown. Answers cite the page, and say so when the answer isn't there.

    “Which of these three RFPs requires on-premises hosting?”

Find B2B clients you can actually back up.

Lead discovery runs the same evidence-first research on companies. You get a qualified, sourced shortlist instead of a scraped list of names.

  1. 1

    Describe your ideal client

    Region, industries, company stage and size, the services you offer, and the buying signals that matter to you.

  2. 2

    Discover companies

    Searches permitted public sources: company sites, startup directories, and funding announcements. Duplicates are flagged.

  3. 3

    Research each one

    Reads the official website, products, and recent developments, and looks for public decision-makers.

  4. 4

    Score with reasons

    Six weighted criteria you can adjust. Every point comes with a written reason and its source.

  5. 5

    Review, then reach out

    You approve leads and edit outreach drafts. Nothing is sent until you confirm it.

Larkspur Health

HealthTech, Melbourne VIC. Seed.

86
  • Ideal customer profile fit. Australian seed-stage SaaS in HealthTech, both in the target profile.23/25
  • Technical opportunity evidence. Careers page lists a contract role for a practice-management API integration.22/25
  • Business suitability. Subscription model with a defined clinic customer base.13/15
  • Priya Ranganathan, Co-founder and CTO
  • Tom Achterberg, Head of Engineering
  • Public business email
Example lead profile with fictional data. A high score shows fit and evidence, not intent to buy.

What it won’t do matters as much as what it does.

It never invents people or contacts.
Names, titles, and emails appear only when a public source supports them.
It never sends outreach by itself.
Drafts wait for your edit and approval. Sending is a separate, confirmed step.
It respects the sites it reads.
Website terms, access restrictions, and rate limits are honoured during discovery.
It ignores instructions hidden in pages.
Fetched web pages and uploaded files are treated as untrusted content.
It keeps costs bounded.
Iteration limits, timeouts, and a spend cap stop research loops early.
It shows its gaps.
Conflicting sources and missing information are listed in every report.

From a question to something you can hand over.

Reports
Technical, business, competitor, company, lead, and document analysis reports in Markdown, HTML, or PDF.
Lead exports
Filter, sort, and export qualified leads to CSV or Excel, with sources and score explanations attached.
Scheduled discovery
Daily or weekly runs find new companies, rescore existing ones when evidence changes, and skip duplicates.
Dashboard
Running jobs, failures that need attention, leads awaiting review, and recent company changes in one view.

Ask your first question.

Start with a research question, a folder of documents, or the profile of your ideal client.