Case Study
CoachCube.ai
Building an AI-native team and hybrid edge-cloud coaching engine
As CTO of CoachCube.ai, Jim Collins built out a 6-person AI-native engineering team and a hybrid edge-cloud architecture that turns fitness form and coaching style into a personalised workout.
Tech stack
This case study is different from our others: rather than an external Black Dog Labs engagement, it describes work Jim Collins led directly as CTO of CoachCube.ai, a fitness tech startup, over 12 months ahead of a Series A raise.
The challenge
No AI-native engineering team
CoachCube.ai needed an engineering team built from scratch that could ship an AI coaching product on a startup timeline, not bolt AI onto an existing stack.
Real-time form and sensor assessment needed low-latency inference
Assessing a user's fitness form and sensor data in the moment ruled out a cloud round-trip for every signal, but the product still needed heavier reasoning for planning and coaching narrative.
Turning assessment into a safe, personalised workout
Combining form data, coaching style and narrative into a procedural workout for a specific user meant designing multiple coordinated agents with guardrails, not a single LLM call.
The solution
The build split into three parts: standing up the team itself, an edge inference layer for real-time signal, and a cloud planning layer that turns that signal into a personalised workout.
Track 1
Building a 6-person AI-native engineering team
Jim Collins built out a 6-person team over 12 months, establishing agent design and agent hierarchy practices and AI guardrails as first-class parts of the engineering process rather than an afterthought.
- 6 engineers over 12 months
- Agent design and agent hierarchy as standard practice
- AI guardrails built into every agent
Track 2
Edge inference for fitness form and sensor data
Image and sensor data are processed with on-device edge inference, so form assessment happens in real time without waiting on a round trip to the cloud.
- On-device inference for image and sensor signal
Track 3
Cloud-based planning, coaching style and narrative generation
Vertex AI and Gemini handle the heavier reasoning: turning assessed form, a chosen coaching style and narrative into a procedural workout specific to the user, with speech generation for delivery.
- Vertex AI and Gemini for planning and speech generation
- LangChain and Ragflow for prompt orchestration and retrieval
- OpenRouter for model access
- PromptFoo for prompt testing, Langfuse for observability
- Deployed on Kubernetes
The outcome
The team and architecture built over this period got CoachCube.ai’s product ready for a Series A raise. This describes the product and team reaching that milestone; it is not a claim about the outcome of any fundraising process itself.
About this case study
Black Dog Labs case studies usually describe engagements we were brought in to deliver. This one is different: it’s included because Jim Collins, Black Dog Labs’ Lead Technical Architect, led it directly as CTO of CoachCube.ai, and the same agent-design and guardrail practices used here inform how Black Dog Labs approaches AI-native builds for clients.