How Obsidian gets the numbers right in adviser meetings, without audio ever leaving its own cloud

Financial advice is inherently number-heavy. Obsidian transcribes meetings between advisers and clients with Reson8 running inside its own environment, giving them accuracy on figures, diarization by speaker, and a setup that is private by architecture.

At a glance

A misheard number isn’t just a typo

Obsidian (obsidianos.com) is one of the fastest-growing platforms in wealth management, combining an AI meeting assistant, CRM, and investment platform so everything an adviser discusses with a client is logged in one system of record.

For Obsidian, accurate and private transcription is non-negotiable. Every adviser-client meeting is unique and full of figures: portfolio values, contribution amounts, fee percentages, dates, account references. If a transcription model hears “40,000” instead of “14,000”, the error feeds into meeting notes, CRM fields, and compliance files. In a highly regulated industry, a mistake means needing to re-check all transcription output by hand.

Generic speech-to-text engines are trained mostly on podcasts and general speech, meaning they fall short on the parts of speech that are complex and highly context-dependent, like numbers.

“Numbers were the thing we tested hardest. If the figures in a meeting note are wrong, nothing else about it matters.”Hasnain Bukhari, CEO, Obsidian

Managing client audio in a regulated environment

As a financial services company, Obsidian promises its customers a platform that is safe from data leakage, which is why the platform is built in-house and hosted entirely on its own infrastructure. Client audio is some of the most sensitive data a firm like Obsidian manages, and routing it to a third-party API on someone else’s servers undercuts that promise.

For Obsidian, the requirement from Reson8 wasn’t just a model that performed better, but a model they could host and run themselves in production. That’s why Obsidian deployed Reson8 inside its own AWS environment, installed via Helm chart on EKS. Meeting recordings are processed in batch through Reson8’s prerecorded endpoint with diarization, so meeting transcripts come back ready to be used in note generation.

Obsidian was previously running their own open-source models, which was painful and time-consuming to maintain. They evaluated Reson8’s EU-hosted API, benchmarking output quality against their existing stack. Four weeks after first test access, the agreement was signed. In-cloud deployment was live days later.

“We put real meeting audio through Reson8 from day one, and the transcription accuracy was immediately better than what we were getting, even compared to other leading providers. Running it inside our own AWS environment made the privacy review easy.”Hasnain Bukhari, CEO, Obsidian

For Obsidian, the model quality was the reason to switch, and the deployment is what made it possible. Reson8 is built to get the words and numbers right, and can be deployed wherever the audio needs to live. The same production models Obsidian uses also run on Reson8’s own EU-based infrastructure. With zero subprocessors, Reson8 is the safe option for platforms with strict data boundaries.

Recorded meetings full of domain-specific vocabulary? Run your real audio through Reson8 and compare word error rates against your current engine. Get in touch or get started at console.reson8.dev.