This is a submission for the Gemma 4 Challenge: Build with Gemma 4
What I Built
The Exact Moment It Clicked
Two weeks ago, when Jess posted about the Gemma 4 challenge, I got stuck in a decision-making loop. I didn't know which idea to build, and I had a few competing options.
Usually, when I think about a new project idea, I don't tell anyone until it is completely done. That is just how I like working. I speak with results, not with plans.
Because of that, I did not really have anyone to brainstorm with. I found myself wishing I had a room full of people I could talk through the decision with, to help me figure out which idea to actually commit to.
Then it suddenly reminded me of Edward de Bono's Six Thinking Hats, which I had read about five years ago. And I thought, damn, I wish I had a local AI system where I could actually run that kind of structured discussion.
Then I stopped...
Whoa, wait a second... Why am I wishing for this? Why don't I just build it RIGHT NOW?
And not just build it, but make it fully local on my own PC. No APIs, no cloud. Just something I can run instantly and talk to like a thinking room inside my machine!
That felt like the idea!
What if I could conjure six of those personas on demand, locally, for free, and let them argue about anything I wanted? And even participate in the discussion when needed?
So I built NeuralHats - a local web app where six AI personas, each running on its own tuned instance of Gemma 4, sit around a virtual debate table and argue about any topic you give them. They follow the canonical order. They actually disagree. The Blue Hat, the chairperson, decides when the debate is over. And when the dust settles, a seventh model, the Facilitator, writes a final report you can save as a PDF.
What it Actually Does
- 🎩 Six tuned personas debate any topic you choose
- 🔄 Up to 5 rounds, with the Blue Hat deciding when to wrap up via a
CONTINUE/STOPtoken - 🧑💼 You can join in, claim one of the hats and contribute your own perspective live
- 📡 Server-Sent Events stream each hat's turn the moment it's ready
- 📄 PDF report synthesised by a dedicated Facilitator model at the end
- 💯 100% local: no API keys, no cloud calls, no telemetry, no internet required after setup
Demo
Check out the video walkthrough:
Code
The Essentials
georgekobaidze
/
neuralhats
Six AI personas debate any topic using Edward de Bono's Six Thinking Hats framework. Powered by Gemma 4 via Ollama. Runs fully local.
NeuralHats
Table of Contents
- About
- Features
- Tech Stack
- Architecture
- Project Structure
- Getting Started
- Configuration
- Security
- How to Contribute
- What's Next
- License
- Acknowledgements
- Author
About
NeuralHats brings Edward de Bono's legendary Six Thinking Hats framework to life through AI. Instead of reading about the method, you experience it. Six distinct AI personas debate any topic you choose, each embodying a different mode of thinking.
Each hat is a fully independent AI model persona powered by Gemma 4 via Ollama, with its own system prompt, voice, and reasoning style:
| Hat | Role | Focus |
|---|---|---|
| ⚪ White | The Analyst | Pure facts, data, and objective information |
| ⚫ Black | The Critic | Risks, flaws, and devil's advocacy |
| 🟢 Green | The Creative | Bold ideas, lateral thinking, alternatives |
| 🔴 Red | The Feeler | Emotions, gut instinct, raw reaction |
| 🟡 Yellow |
To run it yourself:
git clone https://github.com/georgekobaidze/neuralhats.git
cd neuralhats
./setup.sh # or .\setup.ps1 on Windows
./start.sh # or .\start.ps1 on Windows
That's it. The setup script pulls Gemma 4, creates the seven custom models, installs the Python and Node dependencies, and start boots the FastAPI backend and Vite frontend together. You'll be debating at http://localhost:5173 within minutes.
Architecture in one breath
React + Vite + Tailwind v4 ──HTTP/SSE──► FastAPI (Python) ──HTTP──► Ollama ──► Gemma 4
│
└──► SQLite (aiosqlite, ON DELETE CASCADE)
Three layers, zero external services. The frontend is a single-page React app with a virtual debate table. The backend is a small FastAPI server with one main orchestrator and an SSE stream. The AI layer is seven custom Ollama models - six hats plus a Facilitator, all built from the same Gemma 4 base.
Let me walk you through the parts I'm most proud of.
One Base Model, Seven Personalities
Running seven separate copies of Gemma 4 would turn my GPU into lava. Instead, I used Ollama's Modelfile system to create seven lightweight aliases over the same base weights - each with its own temperature, top-p, and system prompt:
# backend/modelfiles/Modelfile.template
FROM {{BASE_MODEL}}
PARAMETER temperature {{TEMPERATURE}}
PARAMETER top_p {{TOP_P}}
PARAMETER num_ctx 8192
The setup script bakes in personality through parameters:
# setup.ps1
$HatParams = @{
white = @{ temp = "0.3"; top_p = "0.9" } # cold facts
black = @{ temp = "0.4"; top_p = "0.9" } # cautious critic
green = @{ temp = "0.9"; top_p = "0.95" } # creative chaos
red = @{ temp = "0.85"; top_p = "0.95" } # raw emotion
yellow = @{ temp = "0.6"; top_p = "0.9" } # warm optimist
blue = @{ temp = "0.3"; top_p = "0.9" } # disciplined chair
facilitator = @{ temp = "0.2"; top_p = "0.9" } # near-deterministic synthesis
}
Red Hat runs hot (0.85) - its job is intuition, gut feelings, vibes. White Hat runs cold (0.3) - its job is facts and only facts. Switching from one to another costs nothing because they all share weights in memory. Personality is just parameters and prompts.
The Blue Hat is a Controller, Not Just a Debater
The Blue Hat is the chairperson. Its prompt forces it to end every response with exactly one of two tokens on its own line:
End your response with exactly one of these two tokens on its own line:
CONTINUE — if meaningful new ground can still be explored
STOP — if consensus has been reached or no new insights are likely
The orchestrator parses that token to decide whether to start another round or end the debate. The LLM's output literally becomes control flow.
# backend/orchestrator.py
def _parse_blue_decision(blue_response: str) -> bool:
"""Return True if debate should CONTINUE, False if it should STOP.
Scans lines in reverse to handle trailing text. Defaults to CONTINUE."""
for line in reversed(blue_response.strip().splitlines()):
token = line.strip().upper()
if token == "CONTINUE":
return True
if token == "STOP":
return False
return True
That tiny function is the heartbeat of the whole loop. Reverse-scanning so trailing whitespace or quote marks don't break parsing. Safe default to CONTINUE because terminating early is worse than running one too many rounds.
The debate loop
Here's the actual orchestrator stripped down. Six hats, in order, up to five rounds, controlled by the Blue Hat's verdict:
HAT_ORDER = [HatColor.WHITE, HatColor.BLACK, HatColor.GREEN,
HatColor.RED, HatColor.YELLOW, HatColor.BLUE]
MAX_ROUNDS = 5
for round_num in range(1, MAX_ROUNDS + 1):
await _push({"type": "round_start", "round": round_num})
for hat in HAT_ORDER:
if hat == user_hat:
content = await _await_user_turn(hat) # human steps in
else:
await _push({"type": "hat_thinking", "hat": hat})
messages = _build_messages(topic, conversation_history,
hat=hat, round_num=round_num)
content = await ollama_client.chat(messages, hat=hat, mode=mode)
conversation_history.append({"hat": hat, "content": content,
"round": round_num, "is_user": is_user})
await _push({"type": "message", "hat": hat, "content": content, ...})
if hat == HatColor.BLUE:
blue_response = content
if not _parse_blue_decision(blue_response) or round_num == MAX_ROUNDS:
await _push({"type": "debate_end", "status": "completed"})
return
That's almost the entire thing. No agent framework, no LangChain, no LangGraph. Just a loop, a queue, and a parsed token. The simplicity is the point.
Real-time streaming with SSE
Waiting 30 seconds for an entire debate to finish before showing anything would be unbearable. So I push each completed hat turn over Server-Sent Events the moment it's ready:
async def event_stream():
while True:
event = await _event_queue.get()
yield event
if event.get("type") in ("debate_end", "error"):
# Hold the connection open briefly so the browser receives the
# final event before the server closes.
await asyncio.sleep(2)
break
The frontend's EventSource reacts in real time, a new chat bubble appears as soon as each hat finishes thinking. Watching it unfold feels like watching a real panel discussion.
🎯 Structured conversation history beats flat transcripts
Earlier on I noticed the hats were ignoring each other. The Yellow Hat would give a generic positive answer that didn't actually respond to the Black Hat's specific risk. That was a context problem, they were getting a flat blob of text and skimming it.
So I restructured the history: separated previous rounds from current round so far, surfaced the most recent Blue Hat direction prominently, and gave each hat per-hat reminders to prevent drift:
_HAT_REMINDERS = {
HatColor.WHITE: (
"REMINDER: Review the conversation history above. Do not repeat any fact, "
"statistic, or metric you have already stated in a previous round. "
"Every sentence must be new information."
),
HatColor.YELLOW: (
"REMINDER: White Hat's data points and Black Hat's identified risks are "
"valuable findings — not just Green Hat's ideas. If you endorsed Green Hat "
"last round, you MUST endorse a different hat this round."
),
HatColor.RED: (
"REMINDER: Pick ONE emotional state for this response and stay in it the "
"whole way through. Do NOT swing between opposite feelings in a single turn."
),
# ... and three more
}
After this change, the debates suddenly felt coherent. Hats started naming each other ("As Black Hat just pointed out..."). The Yellow Hat actually engaged with risks instead of pretending they didn't exist. Same model, same temperatures, just a smarter conversation envelope.
A separate Facilitator
The seventh model, neuralhats-facilitator, runs at temperature 0.2, almost deterministic. It's not in HAT_ORDER. It never debates. Its only two jobs:
- Title generation: when the user types a topic, the Facilitator drafts a short title for the debate
- Final report synthesis: after the Blue Hat votes STOP, the Facilitator reads the entire transcript and writes a neutral, structured summary the user can export as PDF
Splitting it off from the hats keeps the synthesis voice neutral and the temperature low enough to actually be useful as a summary. Mixing those jobs into one of the colored hats would compromise both.
Cascade Deletes
The schema looks like this:
CREATE TABLE rounds (
id TEXT PRIMARY KEY,
debate_id TEXT NOT NULL,
round_number INTEGER NOT NULL,
created_at TEXT NOT NULL,
FOREIGN KEY (debate_id) REFERENCES debates(id) ON DELETE CASCADE
);
CREATE TABLE messages (
id TEXT PRIMARY KEY,
round_id TEXT NOT NULL,
hat TEXT NOT NULL,
content TEXT NOT NULL,
is_user_message INTEGER NOT NULL,
timestamp TEXT NOT NULL,
FOREIGN KEY (round_id) REFERENCES rounds(id) ON DELETE CASCADE
);
ON DELETE CASCADE from messages → rounds → debate means deleting a debate is a single atomic operation. Hundreds of related rows disappear with one DELETE FROM debates WHERE id = ?. No application-level cleanup, no orphaned data, no foot-guns.
How I Used Gemma 4
I went with Gemma 4 E4B as my default base model.
Here's why:
The constraint: it has to be local, and it has to be fast
NeuralHats fires 6–7 model invocations per debate round (one per hat, plus the facilitator for final synthesis). With 5 rounds max, that's up to 31 inference calls in a single debate. If each call takes 30 seconds, that's a 15-minute debate which is pretty unusable.
I needed a model that was:
- Small enough to run smoothly on consumer hardware (laptops, mid-range desktops)
- Fast enough that a hat's response feels like watching someone think, not waiting for a printer
- Capable enough to actually hold a position and engage with arguments, not just produce plausible-sounding mush
Why E4B specifically
The 26B model would have been the safe "capability" choice, clearly better at reasoning. But it still turned out to be too much for the turn-based UX I needed. Each round would take minutes, killing the live-panel feeling.
The E2B (2B) model is lightning fast but it didn't hold its hat persona well enough, under pressure it would drift, lose the role, or repeat itself.
E4B hit the sweet spot. It runs comfortably on a 16 GB VRAM machine, generates a hat response in 3–8 seconds depending on hardware, and is capable enough that with the right system prompt and per-hat parameters it genuinely stays in character. Watching the Red Hat shift emotional tone between rounds, or the Black Hat surface genuinely novel risks each time, that's all E4B.
What Gemma 4 unlocked that nothing else could
Three things, specifically:
1. Native multi-instance personality. Because Ollama lets me create lightweight aliases over the same base weights, I get seven distinct AI personas without seven copies of the weights in RAM. Try that with a hosted API and you're paying for seven independent context windows. With Gemma 4 local, it's free.
2. The Blue Hat's CONTINUE / STOP discipline. Small models often fail at strict format constraints, they want to ramble. Gemma 4 E4B reliably ends every Blue Hat turn with exactly one of those tokens on its own line. Without that reliability, the whole control-flow trick falls apart.
3. The freedom to ship "100% local" as a feature, not a constraint. No API costs, no rate limits, no internet dependency, no privacy concerns about feeding personal dilemmas to a third party. For an app whose entire premise is "let six minds help you think through something you wouldn't want to discuss with anyone else" - that's not a nice-to-have. That's the product.
Summary
NeuralHats started because I was stuck inside my own head and needed another perspective. It turned into a project about how, with the right architecture, a single E4B model can play six different roles convincingly enough to actually help you think.
The Gemma 4 family made that possible, small enough to run on my own machine, smart enough to genuinely disagree with itself, and disciplined enough that a 200-word Blue Hat summary ends with the exact token my orchestrator needs to make a decision.
If you've ever been stuck inside your own head, clone it, run it, give it your problem, and let the hats argue. Worst case, you have a good laugh. Best case, you get unstuck.