Apex Orchestrator The Adaptive Enterprise Intelligence Suite

Why use a sledgehammer when you need a scalpel? Our router analyzes every query and dispatches it to the optimal Gemini model. You get efficiency for simple tasks, performance for complex ones.

Comment

Apex Orchestrator

Apex Orchestrator Banner

Why force a generalist to do a specialist's job?

Our adaptive engine instantly switches gears—using smart routing for speed, vision for diagnostics, and long-context for deep research. You get the perfect tool for every task, ensuring precision without the overhead.

🚀 The Problem

Building enterprise AI agents often forces a compromise:

  • Speed vs. Intelligence: Do you use a fast, cheap model that hallucinates, or a massive, slow model that burns credits?
  • Context vs. Retrieval: RAG is great, but it "shreds" documents, losing the narrative arc essential for legal or financial analysis.
  • Text vs. Reality: Most chatbots are blind and deaf, unable to diagnose physical world problems.

💡 The Solution

Apex Orchestrator is an intelligent middleware that sits between the user and the LLM. It doesn't just pass messages; it thinks about the best way to solve the problem before writing a single line of code.

✨ Key Features

1. Native JSON Smart Router

We replaced fragile prompt engineering with Gemini's Native JSON Schema enforcement. The router deterministically decides:

  • Does this need RAG?
  • Which tools are required?
  • Which model is most cost-effective?
class RouterDecision(BaseModel):
    """Structured output from cheap router; must match Gemini response_schema."""
    needs_rag: bool = Field(..., description="Whether to use RAG retrieval")
    tools_needed: list[str] = Field(default_factory=list)
    model_to_use: str = Field(..., description="e.g. gemini-2.5-flash")
    reason: str = Field(..., description="One-sentence reason")

2. Multimodal "Field Eyes"

Text isn't enough for the real world. Apex allows users to attach Images and Audio directly to the chat.

  • Use Case: A field technician uploads a photo of a cracked engine part and an audio recording of the noise it makes. The agent diagnoses the issue instantly.

3. Long Context "Deep Dive" Mode

Sometimes, RAG isn't enough. We implemented a Long Context Toggle that bypasses vector search entirely.

  • How it works: If the dataset is under 1M tokens (Gemini 1.5 Pro limit), we inject the entire corpus into the context window.
  • Result: Perfect recall for "needle in a haystack" queries across hundreds of documents.

🛠️ Tech Stack

  • Core: Python, FastAPI
  • AI Models: Gemini 1.5 Pro, Gemini 2.5 Flash, Gemini Flash-Lite
  • Validation: Pydantic (Strict JSON Schema)
  • Frontend: React, TypeScript

⚡ Try it out

  1. Ask a simple question: Watch the router pick the "Flash-Lite" model.
  2. Upload a photo: Watch the system switch to Multimodal processing.
  3. Toggle "Long Context": Ask a complex question about 50 PDFs and watch it reason across all of them simultaneously.

Built With

Share this project:

Updates