The output of a one-hour meeting is usually a raw audio file nobody ever opens again. Decisions stay in whoever's head remembers them, action items are never tracked, and owners are never assigned. Most transcription tools on the market are built English-first: accuracy drops sharply on Turkish audio, and the resulting text is unusable without manual correction.

Transify was built for exactly that gap: 95%+ accuracy on Turkish recordings, and not just a transcript — a meeting-intelligence layer that automatically extracts decisions, actions and the people responsible for them.

A user drags in an MP3, MP4, WAV or MOV file, picks a source language, a target language and an output format, and the result lands in their inbox within minutes. Three steps, no technical knowledge required.

What the platform produces:

• Transcript — timestamped raw text with speaker separation
• Meeting summary — decisions, action items and assignees
• Subtitles — in SRT and VTT
• Translation — across 50+ languages
• Content Studio — a blog post, an X thread, a LinkedIn post and a YouTube description from a single upload

In short: one video → six pieces of content. Turning a podcast episode or a webinar recording into a content calendar no longer means opening a separate editor.

Media handling was delegated to FFmpeg. Every uploaded file has its audio channel extracted, normalised and downsampled to a target bitrate before it ever reaches a model. This cut both transcription cost and upload size significantly — sending an entire video file to a model is expensive and pointless.

Long-running work was moved onto an async queue. A one-hour recording does not fit inside the lifetime of an HTTP request. I built the background job layer with Hangfire: the user gets a response as soon as the upload completes, processing continues in the background, and the finished job notifies them by email through the Resend API. The user can close the tab; the job is not lost, and retries and job history are kept.

The model layer was not tied to a single provider. Whisper API handles transcription, GPT-4o handles summarisation and analysis, and MiniMax AI covers additional generation steps. Because each step sits behind its own service, swapping a model does not mean rewriting the whole pipeline.

Data retention was turned into a product feature. Media files are automatically deleted within 24 hours of processing, and the infrastructure is built to comply with KVKK (Turkey's data protection law). Since this is the first question enterprise customers ask, I treated it as a design constraint from day one rather than a policy bolted on later.

Output formats were collapsed into a single conversion layer. STT, SRT, VTT, PDF and Word are all rendered from the same intermediate representation, so adding a new format means writing a new renderer — not touching the pipeline.

The platform is live today: 500+ meetings processed, 50+ languages supported, and measured accuracy above 95% on Turkish audio. The free tier offers a 5-minute guest trial with no signup and 200 minutes per month with an account; the enterprise side adds role-based access, audit logs, team management and a REST API for integrating it into your own application.

For me the real gain was the experience of taking an AI feature out of demo territory and turning it into a billable, observable product with an actual data policy.