Back

Supporting applied AI case study

ThreadScribe Studio

A transcript workspace designed to make AI behavior inspectable, local-first, and recoverable from capture through export.

Implementation

Local release-ready

Focus

Trustworthy AI interaction

Stack

Next.js, TypeScript, SQLite

Year

2026

Inspect source
A 54-second current-build walkthrough of raw transcription, structured cleanup, local persistence, safe export, and recoverable worker failure.

The AI draft never outranks the source.

Most transcript demos end at generated text. ThreadScribe keeps the source, transformation, provider state, and retention behavior in one reviewable workspace.

ThreadScribe timestamped raw transcript with fixture disclosure and deleted temporary audio
Raw timestamped segments preserve the speech record and processing context.
ThreadScribe cleaned transcript with provider mode, confidence, safety, and off-device metadata
Cleaned output remains a labeled draft with model and safety metadata.

45/45

Evaluation before provider theatre.

Five labeled synthetic transcripts test cleanup, notes, actions, empty input, and instruction-like transcript text. The score is a fixture contract result, not a model-quality benchmark.

37

application tests

Transcription jobs, AI schemas, persistence, exports, privacy, and deletion behavior pass the current suite.

2.068s

recorded warm local run

One permitted 8.63-second English clip completed with cached faster-whisper tiny on an Apple M3 Pro.

0

critical or serious axe findings

The automated keyboard and accessibility workflow covers completed, dialog, confirmation, export, and narrow states.

SystemLocal-first architecture

Clear ownership at every boundary.

The browser owns interaction. Server routes own sensitive work. Workers own audio processing. SQLite owns the durable local record.

  1. Browser workspace

    Capture, upload, compare, edit, search, and export remain user-controlled interactions.

  2. Next.js routes

    Server boundaries validate files, orchestrate jobs, route providers, persist sessions, and generate exports.

  3. Transcription worker

    Deterministic fixtures or explicit local Whisper return timestamped segments through one observable job contract.

  4. AI transform runtime

    Clean, Notes, and Actions outputs are schema validated and labeled before the interface renders them.

  5. Local SQLite library

    Sessions, segments, edits, outputs, tags, and safe export records stay searchable on the local machine.

ThreadScribe local Whisper worker failure with retry and retained audio preview

Failure stays truthful and useful.

When an explicitly enabled local Whisper worker is unavailable, ThreadScribe reports the failure, retains the audio preview, offers Retry, deletes temporary audio, and renders no invented transcript.

The previous workspace state remains intact.

Internal paths and raw worker payloads stay hidden.

Provider failures never become fabricated success.

Cancellation follows the same cleanup contract.

EvidenceCase-study record

Local release-ready

What the work proves.

Turn private speech into useful written artifacts without hiding where processing happens, replacing source material silently, or presenting fixture output as live-model intelligence.

Role

  • Independent product direction, UX/UI, Next.js implementation, local persistence, AI evaluation, privacy review, accessibility, and release evidence

User

  • Solo builders who record implementation notes, bug reports, planning thoughts, and short voice memos that need to become durable, searchable work.

Key decisions

  • Kept timestamped raw segments beside transformed drafts so the source remains inspectable.
  • Made fixture, local, and cloud provider modes explicit, with external calls disabled by default.
  • Stored sessions, edits, outputs, tags, and exports in a local SQLite library with verified deletion behavior.
  • Preserved audio previews and prior work across cancellation, worker failure, malformed output, and retry.

Verification

  • Five labeled fixture cases pass 45 of 45 deterministic checks across cleanup, notes, actions, empty input, and instruction-like transcript text.
  • Thirty-seven application tests pass alongside privacy fixtures, keyboard workflow checks, production build, and automated accessibility scans.
  • One permitted 8.63-second clip completed with local faster-whisper in 2.068 seconds warm on the recorded demo machine.

Limitations

  • The public walkthrough uses deterministic fixture transforms. It is not evidence of general model quality.
  • The local Whisper result covers one short English clip on one machine and does not establish production-scale accuracy or latency.
  • Speaker labels and segment confidence remain exploratory; hosted sync, sharing, and compliance certification are outside the project scope.
MoreEarlier / Archived case studies

Earlier work remains available for context and does not represent my current technical scope.