NoveltyAxis: Autonomous Research Idea & Prior Art Discovery Platform
An end-to-end research intelligence platform built in Rust and Next.js that validates novel research ideas against published literature, conducts automated pre-submission paper reviews with line-by-line evidence, and verifies citation integrity without LLM hallucination.
- Live platform
- https://noveltyaxis.com/
- Stack
- Rust, Axum, Python, TypeScript, Next.js, PostgreSQL, Docker, AWS
- Architecture
- Single native Rust binary serving embedded Next.js static export
- Benchmark
- TPAD: 26-item temporally held-out prior-art detection dataset with automatic scoring
- Literature sources
- arXiv, OpenAlex, Semantic Scholar, Crossref
- Core engines
- Prior art mapping, novelty falsification, pre-submission review, citation validation
- Persistence
- PostgreSQL with multi-user workspace isolation and SaaS billing
Background & Research Problem
Academic researchers, doctoral candidates, and R&D engineers frequently spend months developing a hypothesis only to discover late in peer review that closely related prior art already exists, or face desk rejections due to missed literature, flawed reference lists, or unsubstantiated claims. Conventional generative AI tools exacerbate these risks by hallucinating citations, inventing publication volumes, and offering falsely reassuring verdicts that fail under standard academic scrutiny.
NoveltyAxis was designed to provide checkable, evidence-backed research assessments before authors invest significant time or submit papers to peer-reviewed venues. The live platform is available at noveltyaxis.com, delivering verifiable research intelligence without relying on ungrounded language model intuition.
Core Capabilities & Evidence Ledger
The platform delivers three specialized research verification workflows: an autonomous Idea Check that investigates published literature, an automated pre-submission Paper Review, and a deterministic Reference Checker.
The Idea Check workflow formalizes raw hypotheses into structured research questions, constructs multi-hop queries across arXiv, OpenAlex, Semantic Scholar, and Crossref, and rigorously tests for overlap. Unlike standard summarizers, every claim labelled as evidence must quote a retrieved paper verbatim, and the system actively seeks to disprove the user's premise to reveal latent confounding factors and genuine research gaps.
The pre-submission Paper Review evaluates unpublished draft manuscripts. It extracts actionable improvements paired directly with line quotes from the draft, performs dual independent reviewer rubric scoring, and maps prospective publication venues against SCImago Journal Rank (SJR) quartiles (Q1 to unranked).
The standalone Reference Checker validates bibliography entries against public DOIs and registry records in Crossref, OpenAlex, and arXiv. It flags retracted papers, non-existent DOIs, and discrepant author lists deterministically without invoking language models, eliminating hallucination entirely.
Systems Engineering & Full-Stack Implementation
To ensure minimal hosting footprint and low operational latency, the entire backend is compiled into a single self-contained native binary written in Rust. The Rust runtime manages multi-phase asynchronous agent loops, literature rate limits, token budgets, PostgreSQL persistence, and session authentication.
The web interface is built with Next.js and TypeScript, styled with Tailwind CSS, and compiled to an optimized static export served directly by the Rust web server. This unified architecture enables seamless switching between hosted SaaS execution at noveltyaxis.com and offline local CLI investigations using local model endpoints.
The backend was originally written in Python and later rewritten in Rust on Axum.
Agent Design & Evaluation
The investigation agent chooses its next action (search, read, verify or refine) by how much it expects to learn from it, deliberately looks for evidence against the idea, and stops once further steps add little new information.
To measure whether agents can actually recognize already-published ideas, I built TPAD, a benchmark that hides every paper released after a cutoff date and asks the agent to rediscover the held-out work. It ships with a 26-item dataset and an automatic scoring pipeline.