The story behind Conversophy
It began with a conversation with my wife and business partner, Lili, a surprisingly helpful AI analysis of a conversation we'd had together, and the moment she revealed to me why the AI's interpretation and explanation weren't good enough.
Why I built this
I'm James Priest. I've spent a significant part of my working life helping people navigate difficult conversations, improve collaboration, and find better ways to communicate. Much of it I had to learn the hard way myself first, which is exactly why I care about getting it right for the people I help. So when AI became capable of reading long transcripts, I started feeding it real conversations and asking what actually happened, why it happened, and what could be done differently next time.
Some of what it gave back was genuinely useful. It caught things I'd missed while I was inside the conversation, and offered angles I wouldn't have reached on my own.
But it also filled in gaps with things that were never said. It misquoted people, mixed up who said what, and frequently stated an interpretation with far more confidence than the evidence justified.
The standard
The real test came when Lili read one of the reports for a difficult work conversation we'd had. She quickly found places where the analysis had misunderstood what she'd said, or what she'd meant. And once she felt misrepresented, her trust in the rest of it collapsed. At first, that frustrated me, because I could still see real insight in the report. But she'd exposed the standard that actually mattered.
An analysis of a conversation is only useful if the people in it can recognise what happened, and can trust the line between what was said, what can reasonably be inferred, and what simply can't be known.
Could I hand Lili an analysis of one of our conversations and have her say:
That became the Lili Test. It's still the standard that everything in Conversophy is built to meet.
What it took
I first tried to solve it with prompting: a carefully instructed AI assistant, a stricter analytical process, a detailed reference of structures and examples. It got better. It still failed the test.
So I set out to build the system I'd been trying to instruct into existence. What I'd initially imagined as a fairly simple task became a long process of experimentation, revision, and failure. The hard part was never getting AI to say intelligent-sounding things. It was building a system that could understand each line in context: what it meant, what it was doing in the conversation, how it connected to what came before, and how confidently any conclusion could be drawn.
That challenge shaped everything about Conversophy: the forensic analysis, the attention to context, every finding traceable back to the words themselves, and the care to keep observation separate from interpretation.
The goal was never to make AI sound more certain. It was to make the analysis worthy of trust, because the only analysis and feedback that can truly help you grow is one you can believe in.
Philosophy
The analysis serves the person reading it. Every feature is designed to produce insight you can actually use and act on, not output that just looks impressive but fails scrutiny.
Every pattern, finding, and recommendation is linked to something that was actually said in the conversation. Conversophy never asserts what it can't show you.
Some things a conversation simply can't reveal. When evidence is thin or context is missing, Conversophy says so, and marks how confident each finding is, instead of guessing.
The same words mean different things in different relationships. How you talk to your manager is not how you talk to your partner, and Conversophy knows the difference.
Understanding what happened only matters if it changes what happens next. Conversophy turns every finding into specific things to try: skills to build and practices to bring into your next conversation.
They're the most private thing you own. Your data is stored in the EU, never used to train AI, and never sold. Explicit consent for every service, full export, full deletion. Always.
What we believe
The conversations you replay in your head deserve to be understood and learned from, not just endured.
Understanding is only the first step: real change comes from knowing what to do differently, and trying it.
Technology should illuminate, not replace, the human work of understanding each other and communicating well.
Transparency
We use AssemblyAI (Dublin, Ireland) for transcription and Amazon Bedrock (Anthropic Claude, EU-hosted) for analysis, which runs with zero data retention, so your text isn't stored by the AI provider or used to train it. Every service requires your explicit consent, and you can withdraw it at any time.
Transcription
AssemblyAI
Dublin, Ireland
Analysis
Amazon Bedrock
Anthropic Claude · EU · zero retention
Storage
AWS
eu-west-1, Ireland
Become a more skilled, self-aware communicator for the moments, and the people, that matter most.
Analyse a conversation