PeakProfile
Building an AI-powered decision support product that helps mountaineers and guides plan safer expeditions.
Mountaineering has become more accessible, but experience and technical skills do not grow at the same pace.
Social media inspires more people to climb bigger mountains, often framing mountaineering as self-improvement. That can be true, but it can also hide the learning curve. Safe expeditions depend on technical skills, physical readiness, judgement and experience.
Guides often have limited information to understand a client's real ability before planning an expedition. At the same time, climbers piece together information from weather forecasts, guidebooks, forums, accident reports and recent trip reports. The information exists, but it is fragmented, inconsistent and constantly changing.
PeakProfile explores whether AI can help guides and climbers make better decisions by combining structured data, historical knowledge and personal experience into one trustworthy planning assistant.
- —Safety decisions must always remain with the climber and guide.
- —AI recommendations must be explainable and transparent.
- —Information comes from many fragmented and constantly changing sources.
- —Trust is more important than automation.
Because the consequences of poor decisions are real, discovery focused on understanding how experienced mountaineers assess readiness and risk.
- —Conducted 30+ interviews with climbers, mountain guides and SAR volunteers.
- —Studied expedition planning across forums, guidebooks, weather reports, accident reports and recent trip reports.
- —Mapped how route selection changes based on weather, objective hazards, experience and fitness.
- —Compared how experienced and inexperienced climbers evaluate the same objective.
- —Built and tested three AI approaches to understand where users trusted—and distrusted—the recommendations.
Decision support, not decisions
The product never decides whether someone should climb a mountain. It structures information, highlights trade-offs and helps users make informed decisions while keeping accountability with the climber and guide.
Classify before recommending
Instead of asking an LLM to generate advice directly, PeakProfile first classifies the important signals: route difficulty, recent conditions, required technical skills, experience level, fitness and environmental risks. The AI then explains the trade-offs using this structured understanding.
Ground every answer in evidence
Recommendations are linked back to weather reports, route information, accident reports and community knowledge whenever possible, allowing users to understand why the product reached a conclusion.
Route briefing
Combines current conditions, objective hazards, route information and recent community reports into one structured planning view.
Personal readiness profile
Uses climbing history, recent activity, fitness and technical experience to help users understand whether an objective matches their current ability.
Guide matching
Gives guides a richer understanding of a client's experience than a short conversation or simple questionnaire, helping create better matched expeditions.
PeakProfile is currently in development. At this stage the objective is learning rather than scale. These metrics reflect product discovery and early validation.
Building PeakProfile changed how I think about AI products, trust, and decision-making in high-stakes environments.
AI should support judgement, not replace it.
In high-stakes environments, the goal is not to produce confident answers. It is to bring together the right information at the right moment so people can make better decisions.
Trust depends on showing the reasoning.
A recommendation becomes more useful when people can understand what information shaped it, where uncertainty remains, and when human judgement is still required.
Better data matters more than more intelligence.
The quality of an AI product depends heavily on the structure, reliability, and relevance of the information behind it. A sophisticated model cannot compensate for weak inputs.