Pontex Simulation turns a procedure into a scenario an AI can play. The trainee speaks, an AI counterpart answers by voice and stays within procedure, and every response is graded by plain, rule-based code. Scenarios are data, so a new procedure is a new file, not a new program. Everything runs on one computer, with no internet connection.
A trainee calls in a medical evacuation request by voice. KILO 6, an AI dispatcher, answers on the radio, reads the request back with one deliberate mistake, and the trainee has to catch it and correct it. The debrief grades each of the nine lines, shows what was sent against what was correct, and records whether the readback error was caught and how fast.
Measured on one desktop (NVIDIA RTX 5070 Ti, Intel i9-14900K) with Phi-3.5-mini and a Whisper speech model adapted for radio phrasing. The 9-line format follows publicly released U.S. Army doctrine (ATP 4-02.2). Call signs, grid references and casualties are fictional.
Any procedure built from structured spoken exchanges can become a Pontex scenario. Tell us yours.
Cloud AI ruins immersion with network latency and ruins budgets with API costs. Pontex is engineered specifically for high-frequency game loops.
Pontex runs inference in a native C++ backend, outside the managed C# heap, and keeps it off the main thread. Inference is dispatched across worker threads via Unity’s Job System and Burst Compiler, automatically utilizing the player’s CPU, CUDA, or Vulkan hardware.
Forget coroutines. Our LifecycleSystem drives inference natively using Unity's ECS. Token generation is monitored via JobHandle and sampled through Burst-compiled jobs, off the main thread.
Give your AI the ability to trigger in-game events securely. Roslyn Source Generators detect [AITool] attributes at compile-time, emitting direct C# execution code — bypassing System.Reflection entirely for speed and strict IL2CPP compatibility.
Equip NPCs with persistent memory and world lore. The offline Knowledge Baker converts .txt lore into searchable .json vectors. At runtime, the Burst-compiled CalculateCosineSimilarityJob searches thousands of entities in O(n) time.
True multimodal immersion. Convert player microphone input to text instantly via the local Whisper STT engine. Stream LLM responses into the local neural voice engine for dynamic voice acting — synchronized with Unity's AudioSource.
Whether you prefer dragging components or writing unmanaged C#, Pontex adapts to your workflow.
No engineering required. Use the Engine Dashboard to load models. Drop an AgentClient onto any GameObject, assign a Persona, and wire up standard UnityEvents.
Need absolute control? Interface directly with RuntimeNative. Allocate unmanaged pointers, schedule batch inference jobs, and build fully memory-managed pipelines.
using Unity.Entities;
using Unity.Jobs;
using Unity.Collections;
using Pontex.Native;
public partial struct InferenceSystem : ISystem
{
public void OnUpdate(ref SystemState state)
{
// 1. Allocate unmanaged request buffer
var requests = new NativeArray<TokenRequest>(
count, Allocator.TempJob);
// 2. Schedule Native C++ Inference Job
var inferenceJob = new EvaluateLLMJob
{
RuntimePtr = PontexRuntime.GetSharedInstance(),
Requests = requests,
MaxTokens = 128
};
// 3. Dispatch on worker threads
state.Dependency = inferenceJob.ScheduleBatch(
requests.Length, 32, state.Dependency);
}
}Join the teams building real-time games and training simulations on hardware-agnostic, on-device AI. Frame-synchronous. Offline. Yours.