Show HN: I made 6 AI agents debate each other about fantasy football lineups

[machinemusic]

I got tired of choosing between data overload (RotoGrinders) and black-box optimizers that just spit out lineups with no explanation. So I built a brain-inspired multi-agent orchestration framework called Neuron, then used it to create 6 AI analysts that argue about DFS picks in real-time.

Technical stack:

- Neuron framework: Brain-inspired multi-agent orchestration I've been building since March (github.com/ShaliniAnandaPhD/Neuron)

- Landing page for Neuron: https://repo-usher.lovable.app/

- 6 LoRA-tuned models on Vertex AI (trained on T4s for ~$120/month)

- ElevenLabs + OpenAI TTS for real-time voice synthesis Next.js + Cloud Run + Firestore Aggressive caching with Upstash to keep costs at ~$0.04/debate

The agents actually interrupt each other, cite stats, and disagree. Users ask things like "Mahomes or Allen?" and watch them debate it out. One-click export to DraftKings/FanDuel when you're convinced.

Early numbers since July launch:

- 78% 7-day retention - 14 min average session - $230 revenue (users buy voice minutes) - 183 signups for NFL kickoff Sept 5

Demo: fantasyfootballneuron.app

Curious if anyone else has tried multi-agent debates for decision-making in other domains. The Neuron framework handles the tricky parts - interruption management, memory persistence, and conversation coherence.

[machinemusic]

Technical details for those interested:

The Neuron framework handles the orchestration complexity - it's designed around biological neural principles for resilient reasoning and arbitration between agents. This lets the agents maintain distinct personalities while still having coherent debates.

The hardest part was training "personalities" that would actually disagree. Early versions had all agents converging on the same picks. I ended up training each LoRA on different analyst archetypes (contrarian, stats-heavy, narrative-focused, etc.) with curated datasets for each.

For voice streaming, ElevenLabs has great emotion but latency issues, so I mix in OpenAI TTS for snappier responses. Currently building a queueing system to pre-generate common debate segments.

Cost breakdown: ~$0.02 for model inference, ~$0.02 for TTS per debate. At scale this could get expensive, but caching common player comparisons helps a lot. The streaming angle has been interesting - we're going live on Twitch/YouTube during NFL games starting Sept 5. The idea is to have the agents debate in real-time as games unfold, which creates a natural funnel to the product.

Happy to answer questions about the multi-agent orchestration, LoRA training process, or the business side of fantasy sports tools.

[mflehmann]

This is one of the best Show HN feedback loops I've seen in a while.

From a detailed, high-signal suggestion to a live feature link in under 30 minutes is just insane. Kudos to you, @machinemusic, for the build speed, and kudos to @autograd2020 for providing the kind of feedback every founder dreams of.

What's wild to me is how you're speed-running product-market fit right here in the comments. You've clearly found your "first true fan," and their suggestions are a goldmine because they don't just understand the tool, they understand the entire meta-game and the user's mindset. The lesson here for other builders is that finding one person who gets your vision is worth more than 100 lukewarm signups.

This makes me wonder about the business model long-term. The voice minutes are a clever start, but the real moat here feels like the quality of the debate and the persona-tuning you've done. Have you thought about a pro-tier that lets power users like @autograd2020 actually fine-tune their own agents or adjust the underlying data sources? Seems like the logical next step to capture that deep user engagement.

Seriously impressive stuff. Following this project.

[machinemusic]

Love this — and you’re spot on about the moat being the quality of debate + persona tuning. That’s exactly what the Debate Workshop is designed for.

It lets power users fine-tune their agents in two ways:

Quick Setup: upload or link favorite articles, stats, or notes → instantly bias an agent’s reasoning.

Guided Wizard: step-through flow for deeper tuning (data sources, weighting, voice personality).

So yeah, users can actually inject their own edge into the debate. That’s where it gets really interesting: the Workshop turns Neuron from just “six default analysts” into a customizable team that reflects your own process.

Would love your thoughts — does that feel like the right level of control for power users?

[autograd2020]

hey @ machinemusic

Dude, just watched the demo. This is seriously impressive work. You nailed the core problem with most analytics tools: they hide the complexity and the "why" behind a single number.

What you've built here is less of an answer machine and more of a decision-support framework, which is way more valuable.

The Architect agent is the secret sauce. Without it, I could see this becoming a cacophony of conflicting outputs. The way it synthesized the first debate into clear GPP vs. Cash advice was brilliant. It's the component that makes the whole thing actionable.

The persona clash is perfect because it mirrors the exact arguments I have in my own head every Sunday morning. The "barometric pressure" stat from Marcus followed immediately by Big Mike's "are you launching a damn rocket?" was legitimately hilarious and perfectly captured the quant vs. gut-feeling dynamic. And having Zareena, the game theory agent, come in with the contrarian "everyone thinks rain means run, so the leverage is in the passing game" take... that shows you really, really get the DFS meta-game.

To answer your question, yes, this is an incredibly useful pattern. It's a fantastic UI for exploring uncertainty.

As for other personas I'd add, here are a few ideas off the top of my head:

The Vegas Agent: An agent that only speaks in terms of the betting market. It would translate player props, spreads, and totals into implied outcomes. It's a powerful, independent signal that's missing right now. The Scheme Analyst: An agent focused on coaching tendencies and coordinator schemes. "This DC loves to bring a corner blitz on 3rd and long," or "This OC uses motion on 80% of plays." It would add a layer of Xs and Os that's different from the pure player stats. The Beat Writer: An agent that's essentially a real-time RAG on local sports news, press conference transcripts, and beat writer Twitter feeds. It would surface the qualitative "insider" buzz that can often be a leading indicator.

A couple of questions/thoughts:

1. How are you thinking about weighting? As a user, could I dial up the "Quant" and dial down the "Gut Guy" if that fits my process? Or tell the system that I'm in a 150-max entry tournament so it should weigh Zareena's GPP-centric advice more heavily? 2. How do you backtest something like this? I'd be fascinated to see a historical analysis of which agent's advice would have been the most profitable over a season.

Seriously cool build. The application for other domains is obvious, but you've found a perfect product-market fit for this framework right here. Can't wait to see where it goes.

[machinemusic]

Wow, thank you — this comment made my day You absolutely figured out what I was going for: not an “answer machine” but a decision-support framework. Hearing that the Architect agent stood out means a lot — I spent a lot of time tuning it so the debate didn’t devolve into noise.

Love your persona ideas:

Vegas Agent — I’ve already sketched this one; implied totals + props are such a strong independent signal.

Scheme Analyst — great call, coaching tendencies are exactly the kind of qualitative-but-reliable context that would add depth.

Beat Writer RAG — 100%. Local intel is usually the sharpest edge before Vegas moves.

On your questions:

Weighting: That’s on the roadmap — I’m experimenting with letting users “slide” agent influence up/down depending on contest type (e.g. cash vs GPP). Right now weighting is implicit (Architect harmonizes), but user control feels natural.

Backtesting: I’m building a logging pipeline so I can track each agent’s outputs and compare to actual fantasy scoring. Goal is to see which personas would’ve been most profitable over a season, and surface that transparency to users.

The long-term vision is exactly what you called out — uncertainty exploration as a UI pattern. Fantasy is just the wedge.

Really appreciate the thoughtful feedback — curious if you’d personally want the slider/backtest exposed inside the debate UI, or as a separate “analytics view” afterwards?

[autograd2020]

Bro, hell yeah! Stoked you're actually listening to the feedback, that's what's up.

Good question. My take? Keep it clean, dude. The live debate on Sunday morning is the war room, gotta be zero distractions when you're trying to lock in your lineups. The other stuff is for the post-mortem on Tuesday.

Here's how I see it:

Sliders are for Presets. Let me cook up my own 'sauce' in the settings beforehand. Like, I'll make a "Milly Maker" preset where I crank Zareena's GPP brain and the Vegas dude to 11. Then for my cash games, I'll have a "Safe Money" preset that juices the Quant bro. When the debate is live, I just see a little tag that says "Milly Maker Mode: ON". No clutter, just good vibes.

Backtesting is the Scorecard. That’s 100% a separate page. I wanna see that on Tuesday and find out which one of these AI bros actually printed last week. Show me the ROI, the big hits, the bad beats. Gotta know who's got the real alpha so I can tweak my presets for next Sunday.

Dude, this build is insane. Keep cooking. You're onto something huge here.

[machinemusic]

Just shipped your ideas into the app:

Presets = “Milly Maker / Safe Money” modes in settings

Scorecard = post-mortem backtest page

Live here: https://fantasyfootballneuron.app/

— curious what you think!

[autograd2020]

Dude. Just clicked the link. That's insane.

Okay, so you were already way ahead of me with the "Debate Profiles" — that's even slicker than what I was picturing with "Presets." Great minds, bro. It's awesome to see you were already on that wavelength.

Now that I'm looking at it live, the implementation is fire. Love that you show the actual agent weightings right there. Zareena at 90% for the Milly Maker? That's the alpha right there. And the UI to switch between modes is clean as hell.

Quick thought now that I see it: you know what would be a game-changer? Sharable profiles. Let the community cook up their own agent mixes and share them. We could have leaderboards for the most profitable user-created profiles. The "autograd2020 GPP Special" preset, you know?

Seriously, man, the speed is unreal. This is exactly how you build a killer product. Keep shipping.

[machinemusic]

Loved that idea — I went ahead and shipped it. Sharable Profiles + a Leaderboard are now live right inside Profiles.

So now anyone can cook up their own agent mix, share it, and see how it stacks up on the leaderboard week-to-week. Would love for you to check it out →

fantasyfootballneuron.app