austin-starks

How Claude Opus 4.7 rebuilt the data-loading pipeline in my backtesting engine.

Hi HN, a friend submitted this.

Here's the short version:

I was recently fired from my SWE job at Coinbase. When I joined, I fully disclosed a side project I'd been building for 5 years, an algorithmic trading platform called NexusTrade. It was on my resume, discussed in interviews, and disclosed in writing before my start date. HR cleared it.

Months later, Coinbase launched their own AI advisory product (Coinbase Advisor). Their compliance team then retroactively declared my solo project a "material competitor" and gave me 30 days to shut it down, sell it, or resign. When I asked for severance instead, since they were effectively forcing me out over something I disclosed on day zero, I was immediately suspended for "conduct concerns" and fired a week later.

On the technical side: NexusTrade is a MERN stack with a Rust backtesting engine. It has an AI agent (ReAct framework) that can build, backtest, and optimize trading strategies through natural language. Supports stocks, ETFs, crypto, and multi-leg options.

I'm mostly interested in hearing from anyone who's navigated a similar IP/compliance situation as a developer with a side project. What did you do? Happy to answer any questions.

Interesting application of AI outside of development. Unfortunately super expensive right now, but has the potential to be a LOT cheaper with dedicated hardware. What do y'all think?

TL;DR: Use LLMs to analyze unstructured data, use the insights to create trading strategies. Leetcode ez

TL;DR: I am a solo founder building an AI trading platform with more faith in myself than I probably should have. So much faith that I'm willing to publicly embarrass myself by sharing everything with the world.

Here's how you can hold me accountable.

An article describing how to trade with AI the right way.

How I 2x’d the S&P 500 with a single prompt

TL;DR: Treat LLMs as trading strategy engineers, not discretionary traders.

A deep technical document describing how I built a no-code algorithmic trading platform that outperforms the industry standard in terms of performance. If you're interested in performance optimization, system design, or Rust, this is a must-read

TL;DR: BKR and WFRD

better returns, better risk-adjusted returns, and lower drawdown. Is this the future of finance?

As someone who has spent years building algorithmic trading software, this is game-changing. This one platform has ALL of the tools needed to do Wall-Street-grade research and analysis. Have you guys integrated AI into your workflow? Why or why not?

> Vibe trading is a process where you essentially prompt an AI to create profitable trading strategies. AI models act like junior quantitative analysts, and completely automate the work of ideation, financial research, testing, and deployment.

Seems pretty cool. Anybody ever vibe-created a trading strategy? How did it fare in the real world?

The model isn't not SOTA for complex reasoning. It's slower than other models in my testing. And it can't explain the issues with a graph that a 10th grade high school school student can. How is this "PHD-level inteligence"?

TL;DR (Too Lazy; Didn’t Read):

    Google I/O 2025 showcased impressive AI advancements including Imagen 4 and Veo 3 that outperform competitors in image and video generation
    Google demonstrates a strong financial position with Q1 2025 revenue of $90.2B and net income of $34.5B
    It is reasonably valued with a P/E ratio of 19.03 (lower than other FAANG stocks, including Apple)
    Innovative “Other Bets” like Waymo self-driving taxis are already operational and expanding
    Thus, I’m investing some of my NVIDIA profits into GOOGL despite regulatory risks and search advertising dependencies

TL;DR: Rust is powerful, fast, and admired — but it's a terrible first programming language for beginners. I rebuilt my entire algorithmic trading platform in Rust and, while the performance gains were real, the development pain was intense: slow feature iteration, a brutal learning curve, and constant reliance on AI just to get code to compile. For new programmers, focus on languages like TypeScript or Python that let you build and learn faster. Rust’s benefits are real, but they come at a cost that most beginners don’t need to pay.

TL;DR: This article describes a free AI Trading platform and the community of people around it. The platform has over 100 people in it already in one day. If you want to learn about how AI is applied to algorithmic trading, then you should join the community, right now!

TL;DR, this model isn't the major upgrade that people say it is. The model is roughly the same as the previous version, with slower response times, worse reasoning skills, and the same overall coding ability

TL;DR, GPT-4.1 (and Gemini 2.5 Pro) are capable of generating accurate SQL queries for data questions. They do so at a better price AND performance than models like Claude 3.7 Sonnet and Grok 3. This has massive implications for data scientists

"Quasar Alpha" is capable of building complex, profitable algorithmic trading strategies

TL;DR, Even finance isn't safe

These new "Quasar Alpha" and "Optimus Alpha" models performed better than every other large language model, was faster, and free.

The AI race is just getting started.

TL;DR: Google is both better AND cheaper for SQL Query generation

TL;DR we're cooked

TL:DR; I used Claude to create a trading strategy. I then optimized the trading strategy using a genetic algorithm. The result is a strategy that outperforms the market by 3x. It has a better return, better risk-adjusted return, AND a lower drawdown.

Just the Google studio! Thankfully, I'm just doing one-shot testing, so the rate limit doesn't slow me down.

- Gemini 2.5 Pro is great for the price - V3 is AMAZING for an inexpensive, open-source, non-reasoning model - Grok is not that great - O1-Pro was the worse for this task in terms of cost and efficacy.

TL;DR Claude 3.7 Sonnet is the best in terms of pure quality and aesthetics. O1-Pro is the worse (both in terms of quality and cost). DeepSeek V3 and Google's 2.5 Pro are the most budget-friendly options. Grok is grokking.

TL;DR: DeepSeek V3 failed so spectacularly on a brain-dead simple task. 2/10, would not use

TL;DR:

I shared my experience guest-lecturing at Harvard. I talked about how the professor had a surprise challenge question, and that the app failed to answer correctly in front of the students.

I then talked about why it failed, some of my attempts at improving it, and finally how I succeeded. I then talked about how the final solution is significantly better, and that the model is now more than capable of generating these complex rebalancing strategies without any issues at all, even for a list of 100+ stocks.