Argus

Inspiration

We were excited to test out NVIDIA's Nemotron models, and when we realized we could combine the Compound (finance) and Xtract (finding signal in the data) tracks, we put Nemotron to the test in Argus.

What it does

Argus is a financial intelligence website that aggregates key financial information and produces a concise analysis on a stock's investment potential. It fetches the underlying documents — SEC filings, insider-trading forms, market data from Yahoo Finance — and produces a "buy/sell/hold" verdict with a confidence score and a thesis with every claim verified from an official source.

How we built it

Tech Stack:

Frontend: Next.js 16 (App Router), TypeScript, Tailwind CSS v4, shadcn/ui
Data: yahoo-finance2 (quotes, fundamentals, charts, headlines) + SEC EDGAR (XBRL company facts, = submissions, filing text, Form 4s)
AI: NVIDIA Nemotron models (Ultra 550B / Super 120B / Nano) over NVIDIA NIM / OpenRouter; Vercel AI SDK v7 (streaming chat)
Deploy: Vercel (push-to-deploy) — env-driven config, one .env.example that documents the rails

Challenges we ran into

NVIDIA's NIM Platform, the model provider for Nemotron, hit many rate-limits

Accomplishments that we're proud of

Argus can analyze a given stock and produce a buy/hold/sell rating with thesis and confidence score in 10-15 seconds. We worked to get the "thinking time" down as much as possible, from around 70-80s in initial trials.

Argus can also compare two stocks given the fundamental financials in 3-5s. We also reduced this time by a solid amount.

What we learned

Using official sources increases the quality of analysis by a lot - Direct SEC filings is the best public information possible, the company gives detailed financials and growth outlook in an easily-digestible HTML file.
Vercel is.. really good. It used the shadcn library to create a very clean template for Argus in Typescript/Next.js. We were able to improve the analysis of Argus a lot because we didn't have to spend much time on the frontend. Deployment on Vercel was also really easy, I clicked through maybe 3 menus, connected our Git repository, picked a domain name, and it was ready to go. It was our first time using Vercel/v0, but we'll be using it more in the future for sure.
Deepseek is... also really good, and really cheap. I (Joe) used about 50M tokens in planning, implementation, and testing, and the total cost was under a dollar! Using GPT or Claude would have been at least 10x more expensive.

What's next for Argus

Institutional signals: 13F holdings flow - what institutions are loading up or cashing out.
Document-scale grounding: cached filings + retrieval over the full 10-K/10-Q corpus via RAG
More asset classes: Yahoo Finance has data on raw materials, futures, crypto, and more - Argus could deliver insights on those markets given the data.