Moises Prat

After years running large-scale delivery programmes, I kept hitting the same ceiling: executives need cross-programme visibility, but the tools built for execution (Jira, MS Project) are blind to risk at the portfolio level. The result is governance theatre — manual status decks that are stale before they reach the board, and escalations that arrive too late to act on.

I built RoadmapSnap to give PMO leaders the AI signal layer they actually need: dependency health, early risk indicators, and live delivery data — without replacing the tools their teams already use.

• An intelligent governance layer that sits above execution tools, not instead of them
• Dependency-mapped roadmap with visual programme intelligence and real-time risk indicators
• Jira-native integration that pulls live delivery data without replacing existing workflows
• Open-core model: free OSS Lite tier as a distribution engine for the commercial SaaS platform

My motivation was twofold. As someone who follows markets, I wanted a system that could do what I cannot do manually at scale — ingest sentiment, price action, and fundamentals simultaneously and synthesise them into a coherent view. But I also wanted a reference architecture for multi-agent AI that was genuinely production-grade, not a toy demo.

Most multi-agent tutorials stop at "hello world". ProspectAI is my attempt to show what agentic orchestration looks like when you wire it to live data, deploy it with SSE streaming, and publish it as an installable package — the full engineering journey, not just the concept.

• Multi-agent pipeline on CrewAI — 4 agents in sequence, each owning a distinct research stage
• Modal/FastAPI backend streams agent progress in real time over SSE
• Cloudflare Pages frontend renders Bloomberg-inspired dark terminal reports
• Published as an open-source Python package on PyPI — installable in one command
• Documented as a learning resource for AI engineers, with architecture diagrams

Training for three IRONMANs generates a lot of data. I had years of Strava logs, heart rate files, and power data — and none of the available tools could answer the questions that actually matter to a self-coached athlete: am I building fitness or accumulating fatigue? Is this week's poor performance a signal or noise? Should I race this block or rest it?

StrideIQ is my exploration of whether LLM reasoning, combined with deterministic sports science models, can replicate the interpretive thinking of an elite coach. The answer, so far, is a confident yes — which raises broader questions about where AI reasoning adds value over pure analytics that I find genuinely interesting.

• Strict two-layer architecture: deterministic analytics engine + Claude-powered reasoning layer
• Acute/Chronic Training Load (ATL/CTL/TSB) modelling, pace-to-heart-rate ratios, zone distribution
• Automated Strava ingestion stored in Cloudflare R2
• Race predictor covering 5K to Full Ironman using Riegel-formula modelling with triathlon adjustments
• Weekly coaching report engine producing structured JSON with fatigue risk, fitness trend, and recommended focus
• Architected for multi-user SaaS scale