Show HN: I open-sourced an AI sales CRM – and the "experience" problem I learned

  • Posted 3 hours ago by dylanmeyford
  • 2 points
https://github.com/dylanmeyford/radiant-ai-crm-oss
Over the past 12 months I built an AI CRM that changed my life. I funnelled every neuron of sales knowledge I had into it.

Today, I open-sourced the whole thing — code, prompts, agents, orchestration - the whole lot.

What it does:

- Connects to your email and calendar

- Automatically analyses every activity

- Joins, records and analyses every meeting

- Thinks about every deal like a seasoned sales pro and actively drafts the next best action to move the deal forward.

- Consumes your files and playbooks and uses it to draft and inform decisions.

- Sends, Schedules, Drafts and Cancels emails

- Sends, Schedules, Drafts and Cancels meeting invites

- Automatically researches and enriches contacts and deals

- Automatically creates meeting agendas

- Automatically finds new contacts to add to the deal

The output is a set of drafted actions: follow-up emails, meeting agendas, deal risk flags, new stakeholders to add. You wake up, review, approve.

The hard part + key unlock: I built it because I was a founder trying to do EVERYTHING and the sales admin was killing me.

It took me 3 whole attempts to get right. I had to think deeply about what a human sales rep is 'actually' doing, consciously and unconsciously, when running a sales motion.

Key unlock: sales is all about people. The first versions I built were about the emails and calls. Didn't work. BUT, when I orientated the intelligence processing pipeline to first build up a story on each person in the deal, their responsiveness, role, contributions, what they care about, history, and only after roll that into overall deal health and intelligence - that's when it started behaving like a real rep.

Why open-source it?

Honest answer: the math. There are well-funded teams building closed AI CRMs right now. I'm just one dude. If I stay closed, I'm a small fish competing on their terms. Open-sourcing makes me the only one in that category. Funded competitors don't need my code — they have their own infra. What they don't have is an open-source community. A room of genius founders just trying to make the product as great as possible.

The harder problem building Radiant uncovered:

Building this taught me something I can't stop thinking about. Agents are becoming more and more capable. But the gap between an LLM and an actual professional isn't capability - it's Judgement*

An LLM can write a follow-up email. But it doesn't know that a prospect's silence after you mention a competitor isn't them thinking, but actually a deal risk that calls for a battle card and timely landmine email - not a check-in email next week.

I only 'know' that because I learnt those signals the hard way by putting the reps in.

When I built Radiant, what I was really trying to was encode that judgment into a system.

This led me to a side project: a spec called OpenExperts — a structured way for professionals to package their decision heuristics, workflows, and deep domain knowledge into a format any AI agent can consume.

My first attempt will be to take the intelligence pipeline from Radiant and transform it into an 'expert package' - then see if OpenClaw can do it all for me.

The idea is that any agent could 'install' an expert package and immediately be a top tier expert in that field.

Links*

GitHub: https://github.com/dylanmeyford/radiant-ai-crm-oss

OpenExperts spec: https://openexperts.ai

Happy to go deep on the architecture, the orchestration approach, or the OpenExperts idea — curious what people here think about the "encoding expertise" problem.*

0 comments