Asynkit

Problem

Teams that run on voice — loan origination, collections, support escalations, sales calls — share a problem: the quality of a conversation matters enormously, and almost no one can actually see it. QA means a human listening to a fraction of calls and scoring them from memory against a rubric in their head. It doesn’t scale, it isn’t consistent between reviewers, and when a regulator or a customer asks “prove it,” there’s nothing to point to.

The real problem was never transcription — plenty of tools transcribe. It was judgment you could defend: consistent scoring, tied to evidence, across thousands of calls and many languages.

Research

We kept returning to compliance-heavy, voice-driven teams — fintech, call centers, customer success — because for them a bad call isn’t just a bad experience, it’s regulatory risk and lost revenue. That framing decided almost everything downstream: the output couldn’t be a sentiment score or a vibe, it had to be an auditable report a compliance officer would trust.

Decisions

Asynkit is a series of deliberate bets about what to build — and what to leave out:

• Rubrics are templates, not code. A team writes its scoring criteria as a prompt once, and every run uses that same rubric. Scoring stays consistent across reviewers and audio types without us hard-coding any single industry’s rules.
• Every finding cites its evidence. No score exists without a diarized speaker turn, a timestamp, and the exact quote behind it. That’s the line between “trust us” and “verify it in seconds.”
• Compliance gates the whole report. Critical parameters are pass/fail and can fail an entire call — because in these industries, one missed disclosure outweighs a great pitch.
• API-first. The same engine runs as a job API — submit an audio URL, poll for status, fetch the report — so analysis can live inside someone else’s onboarding flow or QA pipeline, not only our dashboard.

Tradeoffs

Template-driven rubrics meant more setup for a team than a one-click “score my call” button — we chose consistency over instant gratification. Insisting on a cited quote for every finding made the pipeline heavier than a raw sentiment pass — we decided defensibility was worth the extra work. And an asynchronous job model (upload → transcribe → diarize → score) makes the product minutes-not-seconds: the right call for an audit artifact, the wrong one for live in-call guidance, which we consciously left for later.

Architecture

At a level a non-engineer can follow: audio arrives — an upload, or a public URL through the API — and is diarized and transcribed into speaker-attributed turns. An AI agent then scores those turns against the team’s rubric template, producing a structured AudioQualityReport: summary score, compliance parameters, weighted section scores, per-tracker observations, timeline, sentiment, actions, and decisions — with every number linked back to a quoted transcript segment.

It runs as an asynchronous job: submit, poll for status, retrieve the report as JSON or an exported document. That job model is exactly what lets one engine serve a single upload in the dashboard and thousands of calls through the API.

Product thinking

The insight the whole product rests on: QA teams don’t need another opinion — they need something they can defend in an audit. So the unit of value isn’t a score, it’s a cited, structured report a compliance officer, a coach, and an auditor can each use without re-listening to the call. Everything else — templates, trackers, compliance gates, the API — exists to make that one artifact trustworthy and repeatable across very different teams.

Business value

One engine serves several markets by swapping the rubric: regulatory disclosures for fintech, hold-and-resolution scripts for call centers, discovery depth and next-step commitment for sales, renewal and churn signals for customer success. The value is the same shape everywhere — take voice, which normally vanishes the moment a call ends, and turn it into an evidence-backed record you can score, audit, and coach from.

What I learned

The hard part of an AI product isn’t the model — it’s defining “good” precisely enough that a machine can score it, and being willing to refuse features so the thing you ship is actually trustworthy. Building Asynkit with a small team made that concrete: the rubric is the product, and every decision to gate, cite, or cut was really a decision about whether someone could stake an audit on the result.