Argus

• 

Inspiration

The generative AI boom flooded research with tools that offer to write papers for you. But in serious academic, legal, and scientific research, ghostwriting isn't just unhelpful—it destroys intellectual integrity.

Long-form research fails in subtle ways that ordinary grammar checkers and LLM writers miss: • Hidden unstated premises and unsupported assertions. • Subtle logical contradictions between distant chapters ( • The "one-shot amnesia" of AI tools that forget previous reviewer critiques when a draft is updated. I have built Argus around a fundamental principle: Be a reasoning partner, not a ghostwriter. I wanted to build an environment that maps the logical structure of an argument, identifies structural vulnerabilities, and guides researchers through rigorous Socratic coaching across multiple revisions.

What it does

Argus is built as an end-to-end cloud-native platform on Google Cloud Platform (GCP):

1. Document Ingestion & Claim Extraction: • Uploaded PDF/DOCX drafts are stored in Google Cloud Storage and parsed asynchronously via Google Cloud Pub/Sub. • Document chunk workflows extract claims, confidence scores, and empirical citations, embedding each claim using Vertex AI Embeddings.
2. Two-Stage Conflict Detection Engine: • Stage 1 (Candidate Generation): High-dimensional Vector Search identifies semantically related claims within and across drafts. • Stage 2 (Verification): Gemini 3.5 on Vertex AI inspects premise incompatibility, determining severity and explanation. (Similarity is not contradiction!)

3. Multi-Agent Socratic Coaching Team: • Powered by the Google Gen AI SDK, three specialized agents collaborate: • Socratic Specialist: Flags unstated assumptions and structural weaknesses. • Evidence Specialist: Pinpoints citation voids and empirical gaps. • Argument Specialist: Audits formal syllogisms and coherence.

4. Revision Intelligence & State Machine: • Argus diffs drafts at the paragraph level and tracks issues through a state machine:

How we built it

Argus is implemented as a cloud-native application using Google Cloud services throughout the processing pipeline.

Application Layer

React frontend FastAPI backend Cloud Run for frontend and API deployment

Data & Event Layer

Cloud Storage for source documents Firestore as the canonical application state Pub/Sub for asynchronous document processing and revision events

AI & Retrieval Layer

Vertex AI / Gemini for claim analysis, conflict verification, and coaching Vertex AI Embeddings for semantic claim representations Agent Retrieval / Vector Search for semantic candidate generation Google ADK for the multi-agent coaching workflow

Intelligence Layer

Argus separates the reasoning pipeline into distinct stages:

Document → Chunks → Claims → Embeddings → Candidate Relationships → Conflict Verification → Argument Graph → Coaching → Revision Tracking

This separation was intentional. Retrieval determines what deserves attention; Gemini determines whether a relationship is meaningful; Firestore maintains the canonical state; and the frontend presents the resulting reasoning structure to the researcher.

Guardrail Layer

A dedicated integrity and content-safety layer sits around the AI workflow. This separates application-specific academic integrity rules from general content safety and allows Argus to remain useful without allowing the AI to take over authorship of the research.

Challenges we ran into

The biggest challenge was to learn the GCP ecosystem. Moving the complete workflow onto Google Cloud introduced practical challenges around asynchronous processing, Cloud Run deployment, Pub/Sub delivery, service authentication, Firestore state management, and integrating retrieval and AI services into one reliable pipeline.

Accomplishments that we're proud of

I am particularly proud that Argus is not simply an LLM wrapped around a document upload interface. It is a stateful reasoning system.

What we learned

The biggest lesson was that building a reasoning system is fundamentally different from building a text-generation application. A useful research assistant needs to understand relationships, state, and change, not just text. I learned that retrieval quality matters, but retrieval alone is not reasoning. A vector database can tell us that two claims are semantically close; it cannot, by itself, establish that the claims are logically incompatible. I also learned that revision history is not merely a version-control feature. It is part of the reasoning model. An issue that disappears from a new draft may have been resolved. An issue whose surrounding paragraph changes but whose underlying assumption remains may still be persistent. An issue that survives multiple revisions deserves a different level of attention. Finally, we learned that guardrails are most effective when they are part of the product architecture rather than simply an afterthought around the model. For Argus, integrity is not merely about blocking bad prompts. It is about designing the entire interaction so that the researcher remains the author and decision-maker.

What's next for Argus

Argus is currently focused on making the reasoning structure of a research document visible and actionable. The next step is to make that reasoning history even richer and to make the intelligence system more scalable than it currently is.

Built With

• fast-api
• firestore
• gemini
• google-cloud
• javascript
• python
• react
• vertex-ai