Inspiration# MetaLLM — The AI That Chooses the Right AI
"Stop guessing which AI to use. MetaLLM figures it out."
💡 Inspiration
I've been building with AI since before it was cool —back when you had to stitch together raw API calls and pray the model understood what you meant. Over time, I watched the AI landscape explode: GPT-4, Claude, Gemini, Grok, LLaMA, Qwen — each brilliant in its own domain, each with its own blind spots.
The problem wasn't access to AI anymore. The problem was choosing.
Every developer I knew — including myself — was jumping between tabs. Paste the prompt into ChatGPT, not satisfied, try Claude, then Gemini. It was fragmented, exhausting, and frankly stupid for a field that's supposed to make life easier.
Then I asked a more interesting question: What if the models could argue with each other?
Not just answer — but debate, challenge, refine. Like having a panel of expert advisors who disagree productively until the best answer surfaces. That was the spark for MetaLLM.
🔨 How I Built It
MetaLLM is a multi-model AI aggregator with three core interaction modes:
1. Debate Mode
Two AI models argue a topic across configurable rounds $(r \in [1, 5])$. Between rounds, a guidance injection layer shapes the discourse. After a minimum of 2 rounds, a neutral judge model delivers a structured verdict with reasoning.
Where $\bigoplus$ is the sequential debate accumulation and $\text{Guidance}_t$ is injected context between rounds.
2. Smart Router
The heart of MetaLLM. Unlike rule-based routers (like OpenRouter), my router is AI-powered — a lightweight Groq model analyzes every incoming prompt and makes three decisions:
- Direct answer (simple/casual queries)
- Route to a single specialist model (by task type)
- Trigger multi-model collaboration or debate
It also performs prompt enhancement before forwarding — rewriting raw user input into optimized prompts. This alone is a significant differentiator.
3. Multi Mode & Direct Mode
Multi Mode sends the same prompt to all models simultaneously for side-by-side comparison. Direct Mode lets users pick one model and switch mid-conversation — with full context preserved.
Model Roster
| Model | Specialized Role |
|---|---|
| Gemini | Real-time web search |
| Nemotron | Data & math |
| GLM 4.5 | Business & strategy |
| Qwen 3 | Multilingual & deep reasoning |
| Kimi K2 | Agentic reasoning & coding |
| Trinity Large | Creative writing |
| LLaMA | Fast & versatile |
Infrastructure
- Backend: node.js + express
- Database: Supabase (managed PostgreSQL)
- Reverse Proxy: Nginx + Cloudflare SSL
- Process Manager: PM2
- Payments: OxaPay (crypto), Lemon Squeezy
- Mobile: Android app (coming )
Monetization Model
Pay as you go Minimum $5 credit to start.
📚 What I Learned
1. Routing is a product, not a feature. I originally thought routing was a small backend detail. It turned out to be the entire value proposition. Getting the AI-powered router right — with prompt enhancement baked in — took more iteration than any other component.
2. Debate Mode reveals model bias structurally. Running systematic debates taught me something subtle: the side that gets argued matters less than the framing of the topic. Some positions are structurally easier to defend because training data skews toward mainstream academic consensus. This is a real research insight, not just a UI quirk.
3. Free tier design is a growth mechanism. Giving real, useful functionality for free (via Groq) while making the queue friction visible — that's not a compromise, it's a conversion funnel. Users feel the difference before they pay.
5. Solo development demands ruthless prioritization. Being a one-person team across blockchain, AI, and full-stack simultaneously means every architectural decision is permanent until it isn't. I learned to build modular, document everything, and ship ugly before shipping never.
🧱 Challenges Faced
Token Counting Inaccuracy
Different providers return token usage in inconsistent formats. My fallback estimator drifts on long conversations. Solved partially with a unified normalization layer — still being refined.
Judge Bias in Debate Mode
The judge model itself carries biases. In feminist vs. traditionalist debates, the feminist position consistently won — regardless of which model argued which side. This isn't a bug, it's a reflection of training data distribution. Documenting and disclosing this transparently became part of the product's honesty.
