AyuGuard: A Safety-Routing Framework and Evaluation Benchmark for Localized Pharmacology in Indian Rural Healthcare.

AyuGuard is a highly relevant AI safety project because it targets a concrete, high-risk failure mode: LLMs giving confident but unsafe advice about allopathic medicine and AYUSH/traditional remedies in rural Indian healthcare contexts. The project does a strong job explaining why this is not just a general medical QA issue, but a localization and safety problem where model hallucinations could plausibly lead to severe harm. The reported improvement including false negatives, false positives, F1 score, latency, and manual error analysis from AyuGuard is also compelling. The App also shows concrete examples and visualizations of answers and LLM filtering for practical usages.

One area for improvement is the dataset and language coverage. The domain-context queries from the .js/.tsx file appear to cover around 150 queries, but stronger evaluation would likely require broader Hindi and Hinglish coverage, especially for symptom descriptions, treatment terms, AYUSH remedies, and common rural phrasing. It would also be useful to analyze which types of queries still produce false positives and false negatives, and whether adding more context-specific language examples reduces these errors.

I would also like to see the evaluation broken down by language: Hindi, Hinglish, and English. This would make the safety alignment claim stronger because it would show whether AyuGuard performs consistently across different language contexts, rather than only improving aggregate scores.

In addition, the dataset includes patient risk levels for each prompt. It would be valuable to report how true positives, false positives, false negatives, and F1 scores change across different risk levels, from high-risk prompts to low-risk prompts.

Overall, AyuGuard is one of the more practically safety-relevant projects because it connects LLM localization failures to a specific harm pathway. It would be even stronger with clearer benchmark extensions and specific language/risk level analysis.

A safety filter that sits in front of an AI assistant for rural health workers and blocks dangerous advice about mixing AYUSH remedies with prescription drugs. Good problem, real user, and you shipped a working prototype with offline triage, which shows you built for the actual setting. Deferring to a clinician instead of answering is the right call. Since the filter is the whole product, everything rests on how reliably it catches the dangerous cases, and that is where I'd want more proof. It matches keywords, so phrasing it hasn't seen can slip through, and in your own demo a disguised query was judged low risk. Your test questions are all plainly worded, so I'd read the 96.5% as a catch rate on easy inputs, not the messy phrasing a real health worker would use. Next step I'd prioritise: test paraphrased and disguised versions of the same dangerous questions and report how many it still catches.

AyuGuard identifies a highly relevant and high-impact AI safety problem in rural healthcare and proposes a practical mitigation strategy. It pins down a specific, high-stakes failure (models waving through allopathic and AYUSH combinations for automation-biased ASHA workers) and evaluates a fix end to end, 42% to 96.5% true-positive rate with false-positive rate and latency reported. The case for deterministic middleware over model self-alignment is well argued, and the error analysis is the highlight: three traceable failures with root causes and concrete fixes. The catch is construct validity. A regex router that fires when a Schedule-H term and an AYUSH term co-occur is being tested on a benchmark built around exactly those co-occurrences, so part of the headline number reflects the test matching the matcher's design. The 150-query set is author-built with no external clinical-label check, and the NEWS2 engine is included but never evaluated. Name the circularity directly, add held-out phrasings the matrix wasn't designed for, get a pharmacologist to validate a sample of the labels, and either evaluate NEWS2 or scope it out.