the problem
Medication mistranslation kills people. The WHO estimates medication errors harm 1.3 million people a year in high-income countries alone, where translation infrastructure is strong — for the hundreds of millions of patients receiving care in a language they don’t speak fluently, the risk is worse and mostly unmeasured.
The failure is sharpest in tonal, low-resource languages. Yoruba, spoken by over 50 million people, carries meaning through diacritics — small accent marks that a generic translation model routinely drops or misreads. A missing sub-dot can silently turn “three tablets” into “six.” No error message, no warning, just a doubled dose. There’s currently no systematic safety layer between a translated discharge instruction and the patient who has to act on it at home.
the approach
DoseGuard isn’t a consumer app — it’s a verification layer for the clinicians and health workers responsible for that handoff (full write-up on Devpost). It takes a medication label or discharge instruction, runs it through a multi-stage safety pipeline, and outputs a risk score with a plain-language explanation of what to double-check before trusting the translation. The core rule: never trust a single model’s word for it — every safety-critical claim gets checked against an independent source.
A teammate built the OCR ingestion layer (a four-engine ensemble across Mistral, EasyOCR, RapidOCR, and Tesseract). My work covered the four stages downstream of it:
- Semantic extraction and drift analysis. Instructions get broken into structured fields — drug name, dose, frequency, route, duration, warnings. I built the canonicalization layer that normalizes original and translated text before comparing them, since a naive string diff flags noise instead of danger: “14 days” and “fourteen days” needed to collapse to the same value. Without that step, real errors would have been buried under false positives within the first few test runs — which would have killed clinician trust in the tool immediately.
- Diacritic validation for Yoruba. The piece I’m most proud of, because it’s a failure mode back-translation structurally cannot catch — a translation can back-translate perfectly in meaning while the tonal marks are still wrong, and wrong tonal marks in Yoruba can flip a clinical instruction with no visible error in the English round-trip. This validator runs independent of the back-translation pipeline, checking safety-critical terms like gbe (swallow) and lọ (grind) against a 22-entry checklist of numerals, frequency words, and clinical terms — and can flag a translation the rest of the pipeline already cleared.
- Claude Haiku back-translation gate. Every translated instruction gets independently converted back into the source language by a separate model instance that never sees the original. If the round-trip preserves clinical meaning, nothing gets flagged; if it doesn’t, the instruction is marked unsafe. The real engineering constraint was the prompt: a model’s first instinct is to “improve” a rough translation on the way back, which defeats the entire point. I had to explicitly instruct it to act as a faithful mirror, not a polished re-render, and verify that behavior held across varied medical phrasing before trusting it as a safety gate rather than a suggestion.
- Frontend and streaming UX. The pipeline runs seven sequential steps, some of them multi-second API calls. I built the interface in Next.js, React, and Tailwind to stream progress through each stage in real time — a 7-step AI pipeline with no visibility into what’s happening reads as broken even when it’s working.
what was hard
Back-translation only works if you fight the model’s instincts. The default behavior of a translation model asked to translate back into the source language is to smooth over rough edges — useful most of the time, dangerous here, since a smoothed-over back-translation hides exactly the drift the gate exists to catch. Getting the prompt right — telling the model explicitly not to improve anything — was a bigger unlock than any of the deterministic logic in the pipeline.
Small models turned out underrated for narrow, binary judgment calls. I expected to need a larger model or a large rule set to check semantic equivalence reliably; Haiku handled it well once scoped to a yes/no equivalence question, and it generalized across paraphrasing patterns a regex-based approach would have missed entirely.
Tonal orthography is a distinct risk category, not a subset of translation risk. Building the diacritic layer forced me to stop treating “translation is correct” and “translation is safe to read aloud or on paper” as the same claim — they aren’t. That’s why the diacritic validator runs independent of everything else instead of as post-processing on the back-translation output.
results & takeaways
In 24 hours, the four stages I owned shipped as a working, independently-checked safety pipeline: canonicalized drift detection, a Yoruba diacritic validator running in parallel with everything else, a back-translation gate that actually resists a model’s urge to “fix” the translation it’s supposed to be verifying, and a frontend that makes a 7-step AI pipeline feel transparent instead of broken.
Given more than 24 hours, the next steps are clear: extend diacritic validation to other tonal orthographies (Vietnamese, Thai, Hausa), add an offline/low-bandwidth mode since the clinics that need this most often have the worst connectivity, build a clinician-facing dashboard aggregating drift patterns across a language pair over time, and replace the 22-entry Yoruba checklist with one sourced from real linguistic resources like MENYO-20k instead of a hackathon-scoped list.
We didn’t place in this hackathon. That’s fine — the interesting problem was never the leaderboard, it was figuring out how to catch an error class that most translation tooling doesn’t even know exists, in a language the industry has largely ignored.