Generative AI development services · LLM, RAG & custom models
We put generative AI inside products that stream, call, record and talk: an assistant that answers from your own content, live captions and summaries, generated media, a model you trained yourself. First feature shipped behind a flag in 2 to 4 weeks, from $6K, with the cost per request measured before you commit. 20 years of real-time video and audio behind it.
Who we build for
WHAT YOU ARE ACTUALLY BUYING
In two to four weeks a generative feature is running in your product: behind a flag, on your data, with technical documentation and a weekly status report alongside. The paperwork describes the feature. It does not replace it.
And one number the rest of this category keeps off the page: what a request costs you in production. In writing, before you commit.
The build decision
You already have users, a codebase and a roadmap, and now someone wants AI in it by next quarter. The demo was easy; production is not. An off-the-shelf AI feature ships this week and gives you someone else's model, someone else's prompt and a per-seat bill. A first in-house AI hire spends months learning evals, retrieval and streaming before the first thing ships. A specialist has already put generative features inside live media products and owns the latency, the cost and the code. Here is the honest trade-off.
Already shipped an AI feature that hallucinates, stalls or costs too much? We pressure-test it and hand you a fix list with numbers. Starting from nothing? We pick the models, ground them in your content and ship one feature you can measure. Need an autonomous agent instead of a generative feature? See AI agents on LiveKit.
What the feature is made of
A generative feature is five layers, not one API call: what goes in, what grounds it, which model runs, what checks the answer, and what the user sees. Here is the pipeline we build, and the latency budget that keeps it usable inside a live call or stream.
Figure 1: Five layers and one latency budget. The pipeline behind a generative feature that holds up inside a live product, with the model layer chosen per task rather than per brand.
We pick the task worth doing first, name the models that can do it, and put a number on cost per request and latency before a line of product code is written. If it is not worth building yet, we say so.
before any product codeYour documents, catalogue, transcripts and recordings become a retrieval index with a chunking strategy, embeddings, reranking, and permission filters, so answers cite your material instead of inventing it.
pgvector or Qdrant, hybrid searchA golden set of real inputs with expected outputs, scored automatically. This is what turns "the demo looked good" into a number you can defend, and it runs on every commit so a prompt change cannot quietly make quality worse.
regression gates in CIStreaming responses, tool calls, fallback between models, moderation, PII redaction, and a hard cost cap. Rolled out to a slice of users, measured against the eval set and the latency budget, then widened.
behind a feature flagCheaper models for easy requests, caching, prompt and retrieval tuning, and observability on latency, cost per request and quality over time. Your team gets the code, the prompts and the eval set.
model routing and cachingThe result: one generative feature live in weeks, grounded in your data, with cost per request and quality both on a dashboard. The hard part is never the demo. It is holding answer quality, latency and unit cost steady once real users arrive. For how we score a feature before trusting it, see evaluating an LLM app in production.
The stack
Three things changed in 2026 and made this worth building now. Frontier models got cheap enough to put in a hot path: a light tier runs at about $0.07 per million input tokens as of 31 August 2026, so classification and extraction cost cents rather than dollars. Real-time speech-to-speech went production, with sub-200ms voice models making a conversational feature feel like a conversation. And open weights caught up, so self-hosting a near-frontier model on your own GPUs is now a cost and data-residency lever rather than a compromise. Named tools below, all verified in August 2026. We pick per task, and we write the choice down with the cost attached.
Model names move every few weeks, which is exactly why the choice belongs in a routing layer rather than hard-coded in your product. We build the seam so swapping a model is a config change and a rerun of the eval set, not a rewrite. For the architecture behind multimodal, real-time systems, see the multimodal agentic AI guide.
What we build
We built StreamFog, an AI effects platform for live streamers and the successor to the discontinued Snap Camera: 300+ AR lenses, 100+ VTube avatars, AI-generated custom face masks, running as a virtual camera into OBS, Zoom and Discord with no re-encoding and about three minutes of setup.
On Franchise Record Pool, Funkmaster Flex's DJ platform, DJs ask in plain speech ("make a playlist with Italian pop from the 90s, BPM 140") across 720,000+ licensed tracks. Built with OpenAI, Whisper and Amazon Polly on a platform doing 3M page views a month.
For Layers, three bracketed photos go in and one finished HDR image comes out, colour-corrected by a custom neural network on IC-Light Stable Diffusion 1.5: 50% better colour accuracy, 60% less image noise, and 30% faster than manual editing.
Live captions, running summaries, action items, chapters and highlight reels produced during and after a meeting or stream, with speaker attribution. Pure transcription on its own is a different job: see speech to text.
An assistant that answers from your documentation, catalogue, course library or call recordings, with citations, per-user permissions, and a "no answer" path when the content does not cover the question. That last part is what makes people trust it.
Draft lessons, assessments, product copy, metadata and B-roll generated at volume, with a human approval step and provenance kept on every asset. We have shipped a GPT-4 course generator used by US higher-education institutions.
Rent vs own
Buying an AI feature is the fast way to say yes to the roadmap. The question a quarter later is who chose the model, whose content grounds it, who can explain a bad answer, and who owns the bill. It is not a question of size: a product with 200 users gets the same pipeline as one with 200,000. Here is the split.
Figure 2: Build vs Buy on value axes only. Product fit and feature ceiling against control and ownership. Custom wins the top-right at any user count, so there is no scale threshold on this chart.
How we work
One clearly scoped feature: grounded in your content, scored against an eval set, shipped behind a flag and measured in your product. You own everything it is made of.
Hallucinations, latency, moderation gaps, or a bill nobody modelled. We measure what it actually does, hand you a prioritised fix list with numbers, then implement it.
Which of your ideas are buildable now, on which models, at what cost per request and what latency. A written plan and a number, including the honest "not yet" where it applies.
Our AI and real-time engineers join yours, work in your repo and your standups, and leave the knowledge behind rather than taking it with them.
Pricing
Fixed-scope starting points. Each is a floor you build up from, and each ends with something running in your product rather than a slide deck. Final scope depends on how many features, how much of your content has to be indexed, whether the model is an API or self-hosted, and your latency target.
How most of our clients work
When you need one number signed off
Model usage is billed by the provider at their rates, with no markup from us, and we forecast it before you commit. For reference, as of 31 August 2026 a light frontier model runs about $0.07 per million input tokens and $0.60 per million output tokens, a mid tier about $2 and $10, transcription about $0.0045 per minute, and realtime audio 32 USD in and 64 USD out per million audio tokens. Self-hosting an open-weight model changes that arithmetic, and we model both paths. If you want a number before you talk to anyone, run an instant estimate.
Free for qualified projects
Before any build, we will scope the feature or pressure-test the one you have. All four are free, all four are written up, and what we find goes into the build. Pick the one that fits where you are.
Competitor analysis, core feature definition, monetization modeling, and a full launch blueprint — delivered within a week. Written by engineers who'll build what they plan.
An independent review of your system's technology choices, structural components, and workload fit — with a plain verdict on what's working, what's a liability, and exactly what to change to reach your goal. Delivered within a week.
A full audit of your code with every issue documented, evidenced, and located — exact file, exact line. Plus a system architecture review and a prioritized fix roadmap. Not a consultant's opinion. A case file. Delivered within a week.
A specialist review of your video or streaming product covering latency, media server architecture, WebRTC, playback reliability, real-time chat, and scalability. Every finding is specific, located, and fixable. Delivered within a week.
Why Fora Soft
Generative AI is new. Shipping it inside video, audio and streaming products under a latency budget is the part most agencies have never done. That part is our day job, and has been since 2005.
Two decades of video calling, streaming and real-time engagement, with AI features in production across them. Not a generalist agency adding AI: the team that already ships the hard half.
A custom diffusion model in production delivering 50% better colour accuracy and 60% less image noise than manual work. Anyone can call an API; fewer can tell you when a fine-tune is the cheaper answer.
Voice in, model, voice out, designed for under a second inside a live session. We budget each hop and pick the model tier and voice engine that fit the number, rather than discovering the lag in production.
Your build ships with a golden set and a regression gate, so answer quality is a number rather than an impression, and a prompt or model swap cannot silently degrade the product.
Per-request cost modelled and capped, frontier API compared against self-hosted open weights, with the forecast handed to you before you commit rather than after the first invoice.
EU AI Act Article 50 disclosure and content marking built in from the start, not bolted on after a legal review. Your code, your prompts, your eval set, your weights, your provider accounts. No lock-in to us.
FAQ
What is generative AI development?
How much does generative AI development cost?
Which model should we use: GPT, Claude, Gemini, or open weights?
How do you stop it from making things up?
Can generative AI run in real time, inside a live call or stream?
Do you train custom models or only call APIs?
Will our data be used to train someone else's model?
What does the EU AI Act mean for our AI feature?
How long until the first feature is live?
Do we own the code, the prompts and the weights?
Keep reading
Multimodal agentic AI in real-time systems
The architecture underneath >LearnGenerative AI for contextual video intelligence
Generation inside a video pipeline >LearnContext engineering for AI agents
Why grounding beats prompting >Neighbouring pages, so you land on the right one: autonomous agents on the LiveKit stack are covered on AI agents on LiveKit, agents that watch video on AI video agents, and phone agents on AI call agents. For speech on its own, see speech to text and text to speech. For AI running on the handset, see on-device and mobile AI. For connecting an existing model or vendor into your systems, see AI integration services. If you want the distinction between an agent and a generative feature spelled out, read agentic AI compared with generative AI.
Tell us the feature, the content it has to answer from, and where it sits in the product. We will name the models, put a cost per request and a latency number on it, and give you a plan in one call.
Specialist software house for video, real-time and AI products. Founded 2005. 50 in-house engineers.
