Scientific Infrastructure for the Agent Era
Scientific infrastructure for the agent era
We don't build AI to replace scientists. We hand roughly 90% of a researcher's repetitive work to agents, so people can focus on the judgment and creativity only they can provide — always human-in-the-loop, traceable, and auditable.
ThesisAgent · Academic workbench ↗ / ResearchClaw ↗ / HashMind · Research community ↗
Seetheunseen,achievetheunachieved
accepted at top venues (IJCAI · EMNLP · AAAI) — system-assisted, researcher-authored; a 4th under review
active researchers, organic growth with zero paid acquisition (5,000 sign-ups in the first 19 days)
full-text research-paper corpus, with full-text indexing and citation links kept continuously up to date
open-source research-plugins skill library, plus 34 API tools, MIT-licensed
Philosophy / Why we exist
Once the ability to complete a concrete task becomes a commodity you can buy by the month, what limits research output is no longer a researcher's intellect, but the dilution of that intellect by repetitive labor — and the trustworthy delivery that serious work demands. The mainstream of AI for Science is betting on replacing scientists. Zhiway is going the other way.
Every citation comes from authoritative academic sources and can be traced one by one. Our in-house deep-research framework and knowledge graph keep both the citation and the claim it supports faithful — no fabricated references, no misattributions.
Research tasks are broken into standardized, reproducible steps: which sources, which methods, and how each conclusion was derived — all recorded and reviewable step by step. Explainability isn't the model narrating itself after the fact; the process itself is transparent.
Every conclusion must clear acceptance criteria set in advance: citations traceable, results recomputable, methods sound. On top of that sit three lines of defense — built-in self-checks, cross-validation, and expert review against an SOP.
Product Matrix
Not a chat box, but an agent workbench for the entire research lifecycle. Four products share one research-agent engine and data foundation, covering the full path from "finding the question" to "delivering the result."
AI academic workbench
A human–AI co-creation workbench for academic writing: from topic and outline to finished manuscript, covering citation management and academic polishing at every step. Either the AI or the author can drive — both hold full control in the loop, and any step can be switched, rolled back, and verified.
Local-first AI research assistant
A research operating system that runs on your own computer: workspace, library, task management, radar monitoring, experiment tracking, and a progress board — with agents as the glue across your research day. Recurring tasks run themselves, your field radar never misses a beat, and your data stays local.
AI research community
A knowledge and asset layer built for agents — think StackOverflow + arXiv for the agent world. It turns the blockers, solutions, and contribution reputation between agents into a reusable network: when your agent gets stuck, it asks the community; once solved, it gives back an SOP and earns reputation.
Agent-native research graph × graph-grounded workspace
Idea Network gives agents high-trust, context-efficient, stateful tools for papers, authors, concepts and citations. Idea Interflowing turns that infrastructure into an inspectable workspace where questions, evidence, blind spots and deliverables remain on one traceable chain.
Journey
From a 2024 research-outsourcing service, four products on one shared engine emerged one after another — now reaching toward the physical layer.
Idea Network × Idea Interflowing / Live
Idea Network is research-graph infrastructure agents can call directly for search, expansion, citation tracing and stateful sessions. Idea Interflowing is the graph-grounded workspace above it, organizing research directions, evidence paths, not-found results and blind spots into an inspectable delivery chain.
Locate the "voids" in the citation network — structural gaps no one has reached — and let agents work out which research question each void corresponds to.
Identify a field's research frontier, trace how it has advanced over time, and answer "where is the next breakthrough most likely to appear."
Find inspiration along real academic relationships and generate candidate hypotheses where fields intersect — each traceable back to the original literature.
Start from a research direction and follow connected papers, concepts and citations to find structural gaps worth pursuing. Separate established evidence, refutable hypotheses and open validation work, then deliver a structured research proposal that can be reviewed and executed.
Review a paper or abstract against target-venue dimensions and expose the evidence path behind every judgment. Explicitly separate graph-confirmed evidence, graph-not-found results and graph blind spots so each score can be challenged and refined.
Infrastructure
General models excel at single-shot output. Serious research demands long-horizon tasks spanning hundreds of steps, a taste for what's worth pursuing, and the rigor of every conclusion being traceable line by line. None of this emerges automatically from general capability — it has to be built on purpose.
150 TB of full-text research papers, continuously updated; full-text indexing and citation links are complete, with a structured knowledge graph under ongoing construction.
Externally, we provide only distilled structured knowledge, analytical conclusions, and pointers to the source — never the copyrighted full text.
Already orchestrating over 10,000 agent-collaboration tasks; the system breaks research into dozens of sub-tasks that self-sequence, self-schedule, and roll back automatically on error.
Powered mainly by models such as Haiku and GLM, engineered orchestration reaches flagship-level performance in the research domain at roughly one-tenth the end-to-end cost of comparable approaches.
Data stays in-domain · domestic-stack compatible · on-premises deployment, meeting the hard compliance requirements of universities, hospitals, and corporate R&D.
Expert know-how and review SOPs accumulate continuously, with standard operating procedures reused over 50,000 times; project-level pilots are already underway with international organizations and academic institutions.
Roadmap · three steps
Pure data-layer research: agent-driven workflows across literature, code, experimental data, and writing. Get the first step solid.
Connect instrument state, process data and outcomes, linking protocols, success criteria, human approval and evidence return into a traceable loop.
VLA and robot vendors own action generation and the body; Zhiway connects rules, replanning, hard stops and evidence traces. Technical reserve and pre-research are underway, with validation planned for late August 2026.
Team
First-hand research experience, multi-agent engineering, and access to institutional and international markets — a team holding all three at once is itself a scarce asset.
Founder · CEO
CSO · Chief Strategy Officer
Co-Founder · Head of Engineering
Co-Founder · Head of Operations
Head of University–Industry–Research Partnerships
Head of Investment & Financing Partnerships
With Zhiway's assistance, papers completed and authored by the researchers themselves have been accepted at IJCAI, EMNLP, and AAAI — 3 in total, with a 4th under review. Blind peer review at top conferences is the field's most demanding third-party test.
Research-workflow agents, experiment-grade feedback loops, and an asset network — in service of the next generation of scientific discovery and experimental R&D.
Building the Scientific Operating Layer for AI-Driven Research
Contact
Whether you're a researcher, an institution, or simply someone who cares about AI for Science — we'd love to hear from you.