Prototyping AI software in a three week discovery sprint
Total Inter Action came to us with a question that could change the way sales training is done across Australia:
Could artificial intelligence simulate real-world sales conversations and deliver live coaching feedback at scale?
We worked with them to test and explore that idea through a rapid AI software prototyping sprint, using GPT-4o, LiveKit, and Pinecone. The goal was to validate feasibility, surface technical and commercial risks, and gather insights that could shape future development without investing in what would be a six-figure build.
Introducing Total Inter Action
Total Inter Action is an Australian leadership and communication training consultancy known for applying the HBDI framework to help sales teams. Led by founder Paul Izbicki, they wanted to scale sales team training through the use of AI simulated roleplay experiences.
Exploring a scalable AI software coaching product
Paul had already experimented with a custom GPT, feeding it PDFs and prompts to run simple text roleplays. Clients loved the concept, but the next step was to determine whether the concept was technically feasible and commercially viable
Paul engaged us to prove (or disprove) the feasibility of the idea before seeking further investor funding.
A time‑boxed AI software development discovery sprint
We proposed a three-week sprint to develop an AI software prototype that could be seen and heard, not just clicked through.
The objectives were:
- Real-time, voice-to-voice roleplay that feels natural
- Multiple personas reflecting different buyer styles
- Structured rubric-based feedback after each session
- A lightweight React Native app for demo purposes
We approached the project like an experiment, starting small, moving fast, and collecting learnings along the way. The brief was to simulate what it might one day feel like to use. That meant making smart technical decisions that prioritised speed, flexibility and feedback.
We chose a tech stack designed for rapid iteration.
Our AI technology stack included:
- React Native & Expo: a fast and flexible mobile app development framework that allowed us to quickly build a front-end interface.
- GPT-4o: OpenAI’s latest multimodal language model, enabling real-time voice interaction with awareness of tone and content.
- LiveKit: an open-source audio and video framework for setting up peer-to-peer voice rooms and coordinating roleplay sessions.
- Pinecone vector DB: a high-performance database for searching and retrieving persona-specific data and scoring rubrics in milliseconds.
- AWS EC2: we hosted the LiveKit agent on EC2 along with a lightweight Express server to handle secure token generation and session management.
While the prototype had limitations, particularly around reliability and model behaviour, it served its purpose as a testbed. It allowed both teams to explore the idea more deeply and understand what would be required to take it further.
The AI software prototype we developed and what we learned in the process
To bring Total Inter Action’s vision to life, we focused on building a simulation of the coaching experience. We built a prototype that could be tested and experienced.
Our process was as follows:
1. Collaborative planning
Working closely with Paul, we mapped out the ideal user experience based on Total Inter Action’s HBDI-aligned training models.
2. Persona design
We crafted simulated buyer profiles based on real-world scenarios. Each had a defined tone, objection style, and behavioural profile to reflect common sales interactions.
3. Rapid app development
Using React Native and Expo, we built a mobile app that allowed users to select a persona, input meeting context, and begin a live roleplay.
4. Live voice integration
GPT-4o was integrated via LiveKit to enable real-time, two-way voice conversations. These weren’t just chatbot scripts, they were dynamic, spoken interactions.
5. Feedback and scoring
After each session, the AI returned structured feedback using Total Inter Action’s own rubric, delivered via JSON.
After wrapping up the prototype, the team reflected on what we’d learned during this AI software development project.
Our AI software development technical insights:
- Prompt engineering is everything. GPT-4o’s default politeness had to be dialled back to create realistic push-back. This was later confirmed as a model quirk by OpenAI.
- Voice triggers are fragile. Short replies like “sure” or “righto” often failed to advance the flow, requiring guard-rails be built.
- Vector search works. Pinecone delivered persona and rubric data in milliseconds, keeping latency low.
- Cost matters. Each ten-minute session used enough tokens to cost $7-15. Enterprise pricing or cheaper models would be needed before scaling. Other infrastructure components like LiveKit, AWS, and Pinecone remained low-cost in prototype mode, but would scale depending on usage. A future build may consider hybrid models that use smaller LLMs or defer voice handling to reduce token consumption.
The final AI software product prototype
Rather than investing heavily upfront, the prototype now functions as a demo from which further feedback and iterations can be worked upon. Importantly, it also highlighted key blockers early, from cost per session and model unpredictability to the need for clearer persona input methods.
“When it works, it works. But getting it to work consistently is the hard part. Without this prototype, I’d be asking for investment based on theory alone.” – Paul Izbicki, Founder and Global CEO, Total Inter Action
By working in this exploratory and lean way, the team avoided what could have been a costly investment. A full-scale build might have run into hundreds of thousands and failed to meet expectations. Instead, they have a functional concept, a set of technical and design learnings to take into the next phase.
For us, this project helped solidify our approach to building AI software. Starting small with focused sprints to uncover critical insights fast.
For more info on our thoughts and approach to AI, check out our whitepaper The Business of AI where we discuss avoiding getting caught up in the hype and explore practical AI software development opportunities.
What’s next?
This AI prototype proved the concept. Future development would focus on improving reliability, reducing latency, and enhancing testing tools. From here, work would explore scalable infrastructure and alternative models to reduce cost without losing the benefits of real-time audio and contextual response.
It showed us exactly where AI needs human-centred design and strong prompt control to deliver consistently.
"Airteam were fantastic to work with, clear, capable, and genuinely collaborative. They helped us turn a complex idea into a working AI software prototype. The process felt structured but flexible, and their technical skill in AI software development really showed." – Paul Izbicki, Founder and Global CEO, Total Inter Action
Got an AI software project idea, but not sure where to begin?
Our AI discovery sprints can help. You can validate, test, and learn, without over investing. Talk to us about your idea and how we might turn it into a working concept you can ensure success with.
Reach out to us at hello@airteam.com.au or via our contact form.
