Rewangit
- Year
- July 15, 2026
- Role
- AI Engineer
- Client
- PT. Mitra Pinasthika Mulia
- Access
- Internal only
- LangChain
- vLLM
- LiteLLM
- Node.js
Internal system — no public link or screenshots.
What follows describes the problem, the approach and the engineering decisions at a level that discloses no proprietary detail.
Rewangit generates documents that follow the company's standard, taking a meeting transcript as its raw material, either produced by Tuturaksara or supplied directly by the user.
What it does
- Transcript to document: takes an unstructured conversation and produces a document in the required internal format.
- Multiple sources: accepts output from the transcription pipeline, or a transcript the user already holds.
- House style enforced: structure, headings and mandatory sections come from the company template rather than from whatever the model produces unprompted.
Approach
The generation step is deliberately constrained. Rather than asking a model to "write the document", the pipeline extracts the required fields from the transcript first, then renders them into the template. That separation means an incomplete transcript produces a document with visible gaps instead of plausible-sounding invention, which matters when the output becomes an internal record.