Founder & CEO
We audited a four-lens retrieval system for a subtle paraphrase-preference bias — and caught the LLM in our own analysis loop writing the conclusion it expected before reading the results, three times. The discipline that fixed it, and four confounds that each shrank our effect.
We built a four-lens fusion retrieval stack, then ran the full 15-way ablation. The cheapest two-lens pair — one dense, one sparse — beat the expensive four-lens stack on every ranking metric, and a 27B model was one of the weakest lenses.
We discovered that HuggingFace Text Embeddings Inference and standard PyTorch produce vectors with cosine similarity as low as 0.30 for the same model.
