I Built a Real-Time Hallucination Prevention System for LLMs Using Computer Vision

LLMs hallucinate. Everyone knows it. Most solutions involve better prompting, retrieval-augmented generation, or fine-tuning. All of these try to fix the problem inside the language model.

What if you used a camera to catch the LLM lying?

That’s SENSE - a real-time framework that takes an LLM’s claims about the world, checks them against live visual input using computer vision, and flags contradictions before they reach the user. Not prompt engineering. Not RAG. A live visual audit loop.

The Core Problem

Imagine a robot or a smart assistant telling you “Hey! There is a red vase on the desk.” The LLM generated that description. But is it actually true? Is there really a red vase? Is it on the desk or somewhere else entirely?

Traditional hallucination mitigation can’t answer this question because it only lives in text space. SENSE answers it by looking at the actual scene.

How It Works

The system has three core pillars:

VisionProbe - The “eyes.” Takes a video frame and a list of LLM claims, runs object detection, and returns what it actually found, with confidence scores and bounding boxes.

LogicGate - The “judge.” Compares the LLM’s claims against the detected objects. If the LLM claimed “red vase” and vision found no red vase above a confidence threshold, it’s flagged as unverified or contradicted.

TemporalTracker - The “memory.” Holds detected objects across frames. This prevents false negatives from transient detection misses so if an object was confidently seen 3 frames ago but briefly occluded, SENSE doesn’t immediately call the LLM a liar.

The main loop looks like this:

llm_claim = ["laptop", "person", "red vase", "mouse", "book", "vase on desk"]

while cap.isOpened():

   ret, frame = cap.read()

   results = probe.probe_batch([frame], llm_claim)

   tracker.update(results)

   buffered_results = tracker.get_buffered_detections(results)

   final_report = gate.audit(llm_claim, buffered_results)

   viz.draw_results(frame, buffered_results, final_report, is_video=True)

Every frame: detect → buffer → audit → visualize. Real-time, on live

video.

Why This Approach Is Different

Most hallucination research lives in the text domain. SENSE introduces a grounding signal from a completely separate modality of vision which the LLM has no ability to fabricate or rationalize around.

This is called symbolic-neural grounding which uses a deterministic symbolic logic (the audit) to constrain probabilistic neural output (the LLM). The LLM can generate whatever it wants. SENSE checks it against physical reality.

The system is also optimized for NVIDIA Blackwell GPUs, meaning the vision pipeline is built to run fast enough for real-time use cases such as robotics, AR assistants, live captioning, surveillance verification.

What the Visualizer Shows

The Visualizer module draws results directly onto the video frame:

• 
  Green boxes: LLM claims confirmed by vision


• 
  Red boxes: Objects detected but not claimed by LLM (model missed something)


• 
  Orange labels: LLM claims with no visual evidence


• 
  FPS counter: Because real-time means nothing if it’s running at 3 FPS

What’s Built vs. What’s Next

The current framework has:

• Real-time video loop with live audit


• Temporal buffering to reduce false negatives


• Threshold-tunable LogicGate (threshold=0.3 by default)


• GPU-accelerated vision probe

What’s coming:

• 
  Spatial reasoning - not just what is in the scene but where. “Vase on desk” requires positional verification, not just object detection.


• 
  Relational claims - handling claims like “the laptop is next to the mouse” which require understanding object relationships


• 
  LLM feedback loop - sending audit results back to the LLM so it can self-correct, completing the loop


• 
  Benchmarking - running against standard hallucination datasets to get quantitative grounding accuracy

Why This Has Research Value

Multimodal hallucination detection, specifically using live visual grounding as an audit mechanism, is a relatively unexplored niche. Most published work focuses on post-hoc text evaluation or retrieval-augmented generation. A real-time vision-based audit layer is architecturally novel and has concrete applications in robotics, autonomous systems, and AR.

This is heading toward an IEEE paper. If you’re working in this space, I’d genuinely like to connect.

Try It

Code is on GitHub: github.com/zosob/SENSE

Requirements: Python, PyTorch, OpenCV, and a GPU helps. Webcam or video file as input.

Drop a comment if you have thoughts on the spatial reasoning problem which is the hardest part and I’m still working through it.