Intent-based personalization
Onboarding captures interests plus a prompting block for depth, familiarity, and relevance spread, giving the generation system real intent rather than a flat list of topics.
Case Study · AI AUDIO
Kaft: A personalized AI news podcast, generated as conversation
Two-host episodes from a listener's interests, rendered on serverless GPU for morning delivery.
The story
What Kaft needed, and what we built.
A news product competes with an overwhelming supply of content. A listener does not need another feed, they need the right stories, at the right depth, in a format that fits their routine. Kaft answers that with a briefing made for one listener. The prototype question was never whether a model could summarize news and synthesize a voice. A dependable product has to capture intent, choose stories, preserve sources, structure dialogue, render long audio, and meet a morning deadline while controlling cost. KUMO designed and built that connective architecture, turning personalization, editorial quality, conversational audio, and operational reliability into one measurable delivery pipeline.
What we delivered
Six areas of production work.
Two content rhythms
Daily Podcasts deliver personalized briefings on current news, while weekly Deep Dives run longer learning episodes with a fresh daily lead-in added without re-rendering the whole episode.
Dialogue-native text-to-speech
Conversational AI voices render through MOSS-TTSD, grouping each Daily Podcast by story and each Deep Dive by narrative act, so both hosts speak inside one coherent segment instead of alternating narration.
Deterministic script serialization
A pure preparation layer maps hosts to consistent speaker labels, normalizes punctuation, expands letter-based acronyms, and keeps story order deterministic before any GPU time is spent.
Swappable TTS provider layer
A provider abstraction submits render jobs, polls status, and returns audio duration, with segment-level retries and a whole-episode rollback path that de-risks migration.
Serverless GPU delivery
Rendering runs on serverless GPU infrastructure with scale-to-zero, warm-capacity planning, per-render cost accounting, and deadline monitoring with alerts, engineered for reliable morning delivery.
Highlights
Project highlights.
Stack
What it was built with.
- React Native →
- Multi-agent LLM
- TTS
- iOS and Android
Services
What else KUMO can build for you.
Custom Software Development
Web, mobile, and platform software, with AI built in.
→ Web & Mobile DevelopmentProduction web and mobile for teams ready to scale.
→ Product Design (UX/UI)UX research, product flows, and UI, built by the same team.
→ AI Product & PlatformProduction AI products and platforms, built to last.
→More work.
View all case studies
FINTECH · CREDIT · INDONESIA FLIN
How KUMO designed and built a mobile financial-health platform for Indonesian users, connecting identity verification, credit-bureau data, privacy controls, payments, and personalised recommendations into one staged journey.
Read the story ->
CONSUMER HEALTH · UK Gluf
How KUMO built Gluf, an AI-powered gluten-free restaurant discovery app for the UK coeliac community. Award-shortlisted.
Read the story ->
HEALTH & FITNESS · MOBILE · US/UK SweatScore
How KUMO built and evolved SweatScore, a live iOS and Android fitness community app, into a points-and-streak engagement system where movement becomes visible progress and social proof.
Read the story ->Tell us what you're solving for.
We'll listen first, ask the right questions, and follow up with a clear proposal.