An AI product that trains, directs and analyzes TikTok Shop livestreams
As Kariisma's product partner, we built a social commerce platform for brands with AI agents.
In this case study
Executive summary
As the product partner of Kariisma, a US live commerce startup, we built a platform that finds and prepares hosts for brands, directs the livestream with the script and the opportunities on screen, and shows the brand what works.
The environment we set up around the agents, with the project's standards and checks, supports more than 30 features, with a 31% lower cost per task and no code breakage in production.
With AI agents writing the code in a flow where the team designs the rules and decides what goes live, features reach production in cycles of about 2 weeks.
The challenge
Kariisma is a social commerce startup that matches each brand with the host of highest affinity to sell live on TikTok. By training and guiding hosts, the platform prepares the livestream environment and offers features to sell more during it.
Our challenge was to come in and build the product along a roadmap with three fronts, each with its own technology. To keep up with the project's speed, with more than 30 features in 5 months, we created a flow that accelerates this build with AI agents, without giving up the rigor that production code demands.
The case in numbers
- 30+
- features delivered in 5 months, in cycles of about 2 weeks
- -31%
- in cost per development task since the agents joined the environment, measured in benchmark
- 0
- code breakages have reached production since the second month; all of them were stopped before entering the main codebase
- 88%
- is the average monthly growth in livestream hours on the platform
- 7x
- more sales per hour than the brand's previous peak, already in its first livestream with Kariisma
Numbers like these start with an immersion in your business and an AI-supported team, before the first line of code. Let’s talk
Inside the technology
We work with a lean team, side by side with Kariisma's product area, and AI agents write much of the code. What makes this safe is the environment around the agent. Every feature starts from a specification written for both people and agents, with the architecture decisions recorded and examples taken from the project itself. The project's standards are applied while the agent writes, and automated checks review every change before it enters the main codebase. Every change arrives ready for production, whether it comes from a person or an agent.
The same care applies to the product's agents. Every new capability of these agents comes in gradually: first in shadow mode, invisible to the host, then enabled through a feature flag, always with an instant kill switch during the livestream. No agent goes live without being evaluated, and it keeps being evaluated afterwards: a language model judges the answers, and adversarial tests confirm that the agent keeps the brand's tone and does not obey instructions planted in the chat.
This environment is what explains the engineering numbers. Since the second month, no code breakage has reached production: all of them were stopped before entering the main codebase. And the cost of each agent task fell 31% in benchmark.
The solution
A live commerce platform that accelerates brand livestreams on TikTok Shop.
AI interview
During selection, an agent runs the interview with the host, by voice, from the questions to the assessment, and says whether the host has the profile brands are looking for
Brand and host matching
An affinity engine matches the brand's ideal customer profile against the platform's hosts, and the list shows who performed best in campaigns
Rehearsal, the training
30-second drills on sales fundamentals and scenarios in a TikTok livestream simulator, with chat; the AI instantly assesses tone of voice, speech, on-camera posture and reaction to the unexpected
Livestream preparation
The brand sets the product queue, the talking points and the reminders of what cannot be missed, and the host arrives at the broadcast with the script ready
Diirector
On a screen beside the host, the agents follow the broadcast and send back their reading in real time: the product queue changes when an item takes off better than expected, alternative paths appear when the livestream goes off course, and the brand's reminders stay in view
Kariismatics
After the show, the analysis tells the brand what the audience asked, what it reacted to and what converted, to plan the next livestream; the host gets the post-show feedback
The results
Livestream hours on the platform grow 88% a month, on average. On Kariisma's side, the three fronts keep evolving, and the two teams operate as one, in the same environment and with the same evaluations. These evaluations now cover nearly 3 times more situations in which an agent could make up a fact, and each one is corrected before it reaches the livestream.
Already in its first livestream with Kariisma, one brand reached 7 times its previous peak in hourly sales. Today, one of the 20 largest brands on TikTok Shop in the United States runs its livestreams on the platform, every day.
In a few weeks, we set up with you the environment and the team to put your startup's product into production. Let’s talk
How we worked
Kariisma came with the product in motion and the founders in charge of it. Before the code, a 3-day immersion with the founders and the investors in the room defined the scope: the journeys of the brand, the host and the livestream, and the roadmap for the product and the go-to-market.
Supported by AI, a single team took on three roles alongside Kariisma. Product led the immersion, mapped the journeys and wrote the specifications for each phase. Product design created the screens and the design system and tested them with real hosts. AI engineering set up the coding agents' environment, the operation of the agents in production and the evaluations, and built features end to end.
This is how we work with startups: we move at their speed, with a lean team of experienced people who design the rules and write the code with AI agents.
Questions this case answers
What is harness engineering?
It is the setup of the environment in which AI agents write code: the standards, the automated checks and the architecture decisions they follow. With it, code written by AI arrives ready for production. At Kariisma, with this environment, no code breakage has reached production since the second month, and the agents' cost per task fell 31% in benchmark.
How long does it take for an AI product to be ready?
It depends on how well it is defined: it can be days or weeks. At Kariisma, a product, design and AI engineering team made the platform available to hosts in 4 weeks, and in about 6 the Diirector was directing a real livestream, with real brands and hosts.
How do you reduce AI agent hallucination in production?
By measuring every agent all the time, with a language model as the evaluator, adversarial tests, reference answers and comparison between models. No agent is free of errors, but the risk falls and becomes visible. At Kariisma, the number of mitigated hallucination scenarios nearly tripled.