Generative AI app development

Generative AI apps that hold up in production

A demo that calls an AI model takes an afternoon. A product that does it thousands of times, stays affordable, recovers from failures and labels its output takes real engineering. We've built three of them, so we know where the time goes.

  • 20+ image, video, voice and music models live in one product
  • Cost previews, usage limits and automatic refunds when a run fails
  • Visible AI labels and IPTC metadata on generated images
Postcrest's multi-shot video planner: AI writes a description and a detailed prompt for each shot before anything is rendered
What we can build

What we can build with generative AI

From one AI feature inside your existing product to a complete AI-first platform. These are the building blocks we've shipped.

Image generation and editing

Text to image, image editing, background removal, upscaling and custom-trained models for a consistent look, with the right model picked per task.

Video generation

Image to video, multi-shot videos planned by AI, talking avatars with lip-sync, captions burned in with ffmpeg and exports in every aspect ratio.

Voice, music and transcription

Text to speech, voice design and cloning, music generation and transcription, integrated with providers such as ElevenLabs and OpenAI.

Text and structured output

Captions, scripts, product descriptions and replies in several languages, with structured output so the result always fits your data model.

Vision and photo analysis

Classify uploads, recognise rooms or products, spot collages or existing AI labels and write alt text, before anything is generated.

Multi-step AI workflows

Chains of models that plan, generate, check and assemble, running as background jobs with live progress in the interface.

Proof, not promises

Generative AI we run in production every day

Postcrest gives creators more than 20 AI models for images, video, voice and music in one workspace, with one credit balance across all of them. Twelve Lumens and Salestreak use image models to stage rooms and restyle product photos, with a vision model checking every upload first.

Calling the model was never the hard part. Everything around it was: queues, webhooks, retries, fallbacks, storage, cost per run, and making sure a failed render never costs the user anything.

  • One interface over many providers, including OpenAI, Google, ElevenLabs, Runway, Fal and Replicate
  • Signed webhook verification and automatic retries for long-running generations
  • Content-filter refusals retried once on a more lenient model; stuck jobs failed and refunded
  • Live progress in the interface while jobs run in the background
  • Generated images labelled visibly and in their metadata (IPTC digital source type)
The same living room rendered as a photorealistic interior
Pencil sketch of a living room layout with handwritten notes
Sketch Render
From sketch to photoreal in the same pipeline (Twelve Lumens). Example images, generated for illustration.
Postcrest's video model picker listing Grok Imagine, Gemini Omni Flash, WAN 3.0, Kling 3.0 and MiniMax with their credit costs
One model picker across providers, each with its cost per run (Postcrest).
AI models live in one product
20+
products running generative AI
3
languages generated natively
5
credits charged for a failed run
0
Who it's for

Who we build AI products for

AI startups

You have a sharp use case for generative AI. We build the product around it: accounts, billing, queues and an experience people come back to.

SaaS companies adding AI features

Your customers ask for AI. We add it to your existing product, with usage limits and costs that fit your pricing.

Media, marketing and content teams

Produce images, video, voice-overs and copy at volume, in your style and in several languages, with review steps where they matter.

E-commerce and real estate platforms

Product photos, staged rooms and listing copy at scale, consistent with your brand and clearly labelled as AI.

How it works

How we build a generative AI product

Models change every few months. The product around them shouldn't have to.

  1. 1

    Prove the output first

    We test candidate models on your real inputs and compare quality, speed and cost per result before building any interface.

  2. 2

    Design the pipeline

    Which model does what, in which order, with which checks. Prompts and structured outputs are versioned like code.

  3. 3

    Make it reliable

    Background jobs, webhooks, retries, fallbacks to a second model and timeouts, so a slow or failing provider never breaks the experience.

  4. 4

    Control the costs

    Usage limits per plan, cost previews, caching and refunds for failed runs, so margins stay predictable as usage grows.

  5. 5

    Label and store safely

    Private storage, signed URLs, metadata stripped from uploads and clear AI disclosure on outputs.

  6. 6

    Swap models as they improve

    Providers sit behind one interface, so moving to a better or cheaper model is a small change, not a rebuild.

AI providers and tooling we work with

We pick the model per task, not per hype cycle, and keep every provider behind one interface so you can switch when a better one ships.

  • OpenAI
  • Google Gemini
  • ElevenLabs
  • Runway
  • Fal
  • Replicate
  • Ruby on Rails
  • Solid Queue
  • ffmpeg
  • libvips
  • Cloudflare R2
  • PostHog
Cost & timeline

What does a generative AI app cost to build and run?

Two budgets matter: building the product, and running the models afterwards. We estimate both. These factors move them most:

Type of media
Text is cheap per run, images cost more and video costs the most. The mix decides your running costs and your pricing.
Number of models and steps
One model call is simple. Pipelines that plan, generate, check and assemble need more engineering and more testing.
Quality bar
How much variation is acceptable? Strict consistency (one style, one character, one house look) takes extra prompt and pipeline work.
Volume and speed
Batch processing, priority queues and live progress add work, but they're what makes a product feel fast at scale.
Compliance
Disclosure labels, metadata, data retention and content filtering, depending on your market and your customers.
FAQ

Questions about building a generative AI app

How much does it cost to build a generative AI app?

It depends on the type of media, the number of models and steps in the pipeline, and how consistent the output must be. After a free intro call you get a written scope and estimate, including an estimate of the running costs per generation.

Which AI models do you use?

Whatever fits the task. We've shipped models from OpenAI, Google, ElevenLabs, Runway, Kling, MiniMax and others, through their own APIs and through Fal and Replicate. We test candidates on your inputs before choosing.

How do you keep AI running costs under control?

With usage limits per plan, a cost preview before each run, caching of repeated work, cheaper models for simple steps and automatic refunds when a run fails. You know the cost per result before you set your prices.

What if an AI provider goes down or changes its model?

Every provider sits behind one interface. Jobs retry with back-off, can fall back to a second model and time out cleanly, and moving to a newer model doesn't require rebuilding the product.

Do you label AI-generated content?

Yes, where it matters. We add visible disclosure labels and IPTC metadata to generated images, which helps with transparency rules such as the EU AI Act and with platform policies.

Can you add AI to our existing product?

Yes. We often start with one AI feature in an existing app, with its own limits and costs, and expand once it has proven its value.

Is our data used to train AI models?

With business APIs such as OpenAI's, data sent through the API isn't used for training by default. We check each provider's terms, keep uploads in private storage and strip personal metadata from files.

Have an AI idea that needs to work for real?

Tell us what you want to generate, for whom and at what volume. You'll get an honest view on models, running costs and what to build first.

martin@balk.dev

Typically a reply within one business day.

Useful to mention

  • What you want to build, in a few sentences
  • Who is going to use it
  • Any deadline or launch date
  • Tools and systems it should connect to