Hi builders,
Over the past few months I’ve been experimenting with an idea:
What would a “flight simulator” for high-stakes conversations look like if it were powered by probabilistic LLM behavior instead of scripted flows?
I built an MVP called TACTIKAI — a simulation layer that allows users to rehearse strategic conversations using dynamically generated AI personas.
What it does
• Generates AI personas informed by publicly available leadership profiles
• Runs probabilistic simulations (not scripted dialogue trees)
• Adapts responses dynamically to user inputs
• Produces structured post-session feedback with coaching signals
Stack
• GPT-4 via API
• Prompt orchestration layer
• Session-state memory management
• Persona conditioning logic
• Custom feedback scoring layer
Core challenges so far
1. Preventing persona drift over longer conversations
2. Maintaining variability without losing coherence
3. Avoiding repetitive conversational loops
4. Balancing token cost with multi-turn realism
I’m especially curious how others are handling persona stability across long sessions without over-constraining the model.
Are you relying more on:
• Strong system prompts?
• Memory compression?
• Structured state injection?
• Tool-based role anchoring?
Would genuinely appreciate technical feedback or architectural suggestions.
If anyone wants context on what I built:
- Avg latency per turn: ~2–4s
Kindest regards to all
Santiago
Founder