Machine repair generates records that rarely get written. The form that should capture root cause, corrective action, idle time, and parts used is filled in later, from memory, or not at all, because the next machine is already down. The organization then repairs the same faults repeatedly while the diagnostic knowledge stays in the heads of whoever was on shift. MendLog is a bet that the medium is the problem: forty years of skipped paper forms suggest the form itself does not survive a workshop, and speech does. The technician records a voice memo in Sinhala or English, takes a few photos, and moves on. MendLog transcribes the memo, extracts the structured fields the form would have collected, and files the job.
Retrieval is the other half
A repair archive is useful only if it can be queried. Jobs are semantically searchable by text, photo, or error code through pgvector. The target case is concrete: a technician facing a fault code on an unfamiliar machine should be able to retrieve how a colleague diagnosed the same fault months earlier. That query is the point at which the captured records pay back the cost of capturing them.
Stack
The stack is chosen for the deployment environment rather than for a demo. React Native with Expo, Android first, because the target device is a mid-range Android phone; that constraint makes memory efficiency and offline tolerance design requirements rather than optimizations. Supabase provides Postgres, auth, storage, and the vector search in one service, which is the right amount of backend for a solo project at this stage. OpenAI handles transcription and structured extraction; transcription quality on spoken Sinhala is unverified, and I intend to test it before trusting it.
Status
MendLog is in early development and nothing is wired up yet. The repository contains a README, a license, and a plan; no pipeline runs. Every paragraph above describes intent, and it is better to state that plainly than to write around it.
The next milestone is the decisive one: record a memo on a low-cost Android phone and get a correctly structured job record out the other end, in both languages. That result either validates the core bet or ends the project, and either outcome will be documented.