↓ Skip to main content
Background Image

HaloscoreAI

·133 words·1 min
Inference-Engineering Startup Hallucination Compliance Safety
Manoj
Author
Manoj
ML Engineer @ 7-Eleven
Table of Contents

Pitch
#

HaloscoreAI scores hallucination risk during inference by measuring how much quantized and higher-precision model paths disagree.

The product does not claim to prove truth. It identifies fragile tokens and responses that deserve citation checks, retrieval repair, or human review.

Product Output
#

  • Response-level risk score.
  • Token-level heatmap.
  • Drift report by model/version/precision.
  • Compliance logs for regulated workflows.

Architecture
#

flowchart LR
  Response[Generated response] --> Quant[Production quantized logits]
  Response --> Shadow[Sampled FP16 shadow logits]
  Quant --> Divergence[Distribution divergence]
  Shadow --> Divergence
  Divergence --> Risk[Risk score]
  Risk --> Action[Accept / cite / escalate]

Customer
#

Healthcare AI, legal AI, financial AI, and enterprise systems where unsupported claims create liability.

Risks
#

  • Divergence may not correlate strongly enough with factuality.
  • Shadow precision checks add cost.
  • Buyers may ask for formal guarantees the signal cannot provide.