The problem iSuara-守语 solves

Malaysia has 44,000 Deaf citizens and ~60 certified interpreters at RM150/hour with three days' notice. The absence of two-way communication compounds across every institution: 85% of Deaf students face severe academic isolation, Deaf patients are 7× more likely to receive a dangerous misdiagnosis, and only 0.3% of PwDs hold civil service roles against a 1% quota.

Existing tools do not close this. JARI/MyBIM signs to the Deaf; glove and webcam systems read from them. Each solves one direction, so neither produces a conversation — and all require connectivity, with gloves adding hardware cost. None handle BIM's non-linear Topic-Comment grammar, so recognition alone yields disjointed keywords rather than meaning.

Challenges we ran into

1. A bug that returned HTTP 200 and looked fine
   Our avatar's BIM grammar endpoint has a rule-based fallback for when the router is unreachable. During integration it started returning perfectly plausible output, correct tokens, sensible reasoning, 200 OK but nothing in the logs, nothing in the UI.

It had been silently falling back for hours. A keyword mismatch (model= vs model_id=) was raising a TypeError inside a broad except, which swallowed it and served the fallback instead.

We only caught it because the response's model field said "BIM Linguistic Engine" instead of a Gonka model id. We learn that a fallback that's indistinguishable from success is worse than a crash. We now assert on the model field and the presence of a Gonka Request ID if the router didn't answer, we want to know.