Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle Submission
This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle
What I Built
QuantCoach — a voice-first alpha research copilot for quant researchers grinding through WorldQuant BRAIN.
My friend was stuck in a loop every quant researcher knows too well: sketch an alpha idea on paper → try to write it as a BRAIN Fast Expression → get stuck on the math or operator syntax → lose momentum. The next day they'd forget why the previous attempt failed. Their research notebook was a graveyard of half-finished ideas with no thread connecting them.
QuantCoach is the tool they'd pay for to fix this.
It's a voice-first workspace where a quant researcher can:
- Describe an alpha intuition out loud ("combine improving margins with a short-term reversal filter")
- Get back a candidate BRAIN-style Fast Expression plus the math behind it
- Hear the explanation read aloud so they keep their hands on the keyboard
- Build a permanent memory of which factor families worked and which ones burned them
- Ask follow-up DSA/math questions in the same session without breaking flow
Who pays for this: Quant researchers, retail quants, quant students, and anyone doing systematic alpha research.
Why they'd pay: Serious quant researchers already spend $100–500+/month on data subscriptions, courses, and tooling. A $29/month IDE-style copilot that remembers your research history, generates Fast Expression drafts, and warns you about lookahead leakage before you waste a week — that's a rounding error against the cost of one wasted research cycle.
It's a thinking partner for alpha ideation and quantitative learning. Not a black box promising returns. Not a signal service. An IDE for alpha ideas.
Demo
⚠️ Hosted on Render's free tier — it may take 30–60 seconds to wake up on the first visit. Thanks for your patience.
Code
The repo includes:
-
utils/prompts.py— the specialized BRAIN Fast Expression system prompt (inspectable, forkable) -
.env.example— every key you need to run it locally -
README.md— quickstart, architecture diagram, and pricing rationale
How I Built It
QuantCoach is built with open-source AI at its core, using a modular stack designed for privacy and speed — both of which are non-negotiable for quant work.
The voice-first pipeline
Voice → ElevenLabs STT → Backboard Memory → Groq LLM
→ Fast Expression + reasoning → Backboard save
→ ElevenLabs TTS → Voice Out
The stack
| Layer | Technology | Why |
|---|---|---|
| Inference | Groq + GPT-OSS 120B (open-weight) | Ultra-low-latency. Fast enough that the voice loop feels conversational. |
| Memory / RAG | Backboard | Every alpha idea, every DSA question, every failure — stored so the coach remembers your research history. |
| Voice In | ElevenLabs Scribe STT | Multilingual speech-to-text. Hands-free is the whole point. |
| Voice Out | ElevenLabs TTS | Reads back explanations and BRAIN expressions. Keeps hands on keyboard. |
| Persistence | MongoDB Atlas (optional) | Powers weak-spot tracker and progress dashboard. Falls back to in-memory if not configured. |
| Frontend | Streamlit | Warm almond/crimson "editorial quant" aesthetic with custom Lucide SVG icons. |
The alpha generator's system prompt does four things
- Translates natural language intuition into BRAIN Fast Expression syntax
- Calls out common pitfalls — lookahead leakage, turnover drag, overfitting across lookback windows
- Warns when an idea is likely correlated with known industry factors
- Refuses to promise performance — it's an ideation tool, not a signal service
The prompts live in utils/prompts.py. Anyone can inspect exactly how the coach reasons about alphas.
Why Does Open Innovation Matter?
Open-source AI wasn't a checkbox here. It was the entire reason this project is safe for a quant researcher to use — and safe enough that they'd pay for it.
1. Alpha ideas are proprietary. This is the whole game.
My friend is developing factor hypotheses that could be worth real money. Sending those prompts to a closed API means they enter someone else's training pipeline. Running open-weight models through Groq and storing memory in Backboard means the research context stays private.
For anyone doing serious quant work, this is non-negotiable. No serious researcher is going to paste their alpha thesis into a closed API just to get syntax help. The privacy angle is not a nice-to-have — it's the reason the product can exist at all.
2. Model portability is a research edge.
The alpha landscape moves fast. When a better open-weight model drops, they swap GROQ_MODEL_NAME in .env and the whole system upgrades. No rewrite. No negotiation with a closed vendor's roadmap. No waiting for someone else's product decision.
For a side hustle, that's also a moat. I'm not locked into one vendor's pricing or terms.
3. Zero cost to the learner.
Quant education is already gated by expensive data subscriptions and paid courses. The tooling shouldn't add to that. The entire QuantCoach stack runs on free tiers, making it accessible to anyone with curiosity and a laptop.
That said — the business works because the costs are near-zero at small scale. Once a researcher finds value, $29/month is trivially affordable against what they're already spending.
4. Transparency matters for a coach.
Because the model is open-weight and the prompts are in the repo, anyone can inspect exactly how the coach reasons about alphas. With a closed API, that reasoning is a black box. For a tool that's teaching someone quant thinking, black-box reasoning is unacceptable.
5. The side-hustle economics work because open AI is cheap.
Closed inference APIs would make this $29/month product unprofitable at any real usage volume. Open-weight models on Groq make the unit economics work — which is exactly why this can be a sustainable side hustle instead of a hobby that dies when the bill arrives.
My First User
My original user is the person this was built for — a quant researcher grinding through WorldQuant BRAIN. Here's what happened when they tried it:
- Used it for: 6 alpha ideation sessions over 4 days
- What they said: "The first time it read back a Fast Expression with the math explained out loud, I actually understood why my previous attempt failed. That's never happened before."
- Did they pay? Not yet — it's still on free tier. But when I asked if they'd pay $29/month for it, they said "immediately, yes — I spend more than that on coffee in a week and this saves me hours."
-
What they asked for next:
- "Can it track which factor families I've already tried so I stop re-doing the same idea?"
- "Can it warn me when two alphas are probably correlated?"
- "Can I export my session history as a research journal?"
Those three asks are on the roadmap. The fact that they came up unprompted — from a real user, mid-session — is the strongest signal that this is a product, not a demo.
Would they pay? Yes. Would a quant student pay? Yes, at a lower tier. Would a small fund pay? Yes, at a team tier. The pricing tiers are already sketched in the README.
Prize Categories
I'm entering the following partner categories, each of which is genuinely load-bearing in the project:
Featured categories
- Best Use of Render — Frontend and AI runtime hosted on Render
- Best Use of Groq — Open-weight model inference for real-time alpha ideation
Partner categories
- Best Use of Backboard — Memory & RAG powering the persistent research context. Every alpha idea, every failure, every DSA question is stored so the coach remembers the researcher's history.
- Best Use of ElevenLabs — Full voice loop (Scribe STT + multilingual TTS) for hands-free alpha research. This is what makes the tool usable without breaking keyboard flow.
- Best Use of MongoDB Atlas — Optional persistence layer for the weak-spot tracker and progress dashboard. Falls back to in-memory storage gracefully if not configured.
Top comments (0)