-
-
Demotivator AI
Inspiration
Most hackathon projects aim to motivate, inspire, or “change the world.” We decided to do the opposite. We were inspired by: Toxic startup culture - where every half-baked project is hyped as “world-changing.” Motivational fluff -99% of AI “life coaches” are just ChatGPT with pom-poms. Our own frustration - sometimes the most useful feedback is just: 𝑌𝑜𝑢𝑟 𝐼𝑑𝑒𝑎 = Garbage That painful hit of honesty is what Demotivator AI delivers.
What It Does
Demotivator AI roasts you, crushes your ego, and exposes every flaw in your ideas.No pep talks.No sugarcoating.No fake positivity.If your startup is a donkey dressed as a unicorn, we’ll call it a donkey. Brutally. Honestly. Sarcastically.
How We Built It
Model Backbone: gpt-oss-20b, one of the largest open models. Quantization: Tried mxfp4, AWQ, even considered GPTQ to squeeze the beast into VRAM. Serving: Used vLLM for OpenAI-compatible APIs. Backend: FastAPI + Uvicorn to handle requests. Frontend: Minimalistic terminal/optional UI — because the roast, not the interface, is what hurts. Code snippet of roast delivery:
@app.post("/roast") async def roast_user(input_text: str): response = openai.ChatCompletion.create( model="gpt-oss-20b", messages=[ {"role": "system", "content": "You are a brutal demotivator AI."}, {"role": "user", "content": input_text} ] ) return {"roast": response["choices"][0]["message"]["content"]}
Challenges We Ran Into
CUDA Out of Memory
torch.OutOfMemoryError: Tried to allocate 1.08 GiB
Translation: buy a bigger GPU.
Disk Storage Hell Downloaded the 20B model ~7 times. Nearly nuked our C: drive. Quantization Myths Papers said “20B fits on 8GB VRAM.” Reality said “LOL, no.” Tone Balance Making it funny brutal, not depressing brutal.
Accomplishments That We’re Proud Of
Actually got gpt-oss-20b running on our setup (after multiple GPU meltdowns). Built a fast roast pipeline with vLLM + FastAPI.Managed to turn sarcasm into a product.Accepted that sometimes deleting 50GB of failed checkpoints is progress.
What We Learned
Big models are cruel — in VRAM, disk, and words. WSL eats space like a black hole — learned to nuke cache directories properly. Hackathon coding = therapy — half the time debugging CUDA errors, the other half laughing at roasts. Sarcasm is addictive — once you build a roast bot, you never want to build a cheerleader again.
What’s Next for Demotivator AI
Fine-tuning on startup pitch datasets — so it can roast like a seasoned VC. Add voice mode — because being roasted out loud hurts more. Roast-as-a-service (RaaS) — integrate into pitch practice, career coaching, even dating apps. Demotivation Leaderboard — rank the most destroyed egos at the hackathon.
Closing Thought
Most AI projects pretend to “change the world.” Ours doesn’t. Demotivator AI exists for one purpose:
To remind you that your idea probably sucks. But hey — at least now you know why.
Log in or sign up for Devpost to join the conversation.