alkemy

I was a consultant for three years, and the hardest part was never the work. It was business development.

Sales in professional services comes down to two things: the right moment and the right person. To find the moment, you dig through filings, news, and LinkedIn looking for a reason to call a buyer. A good partner can hold that loop in their head for three or four accounts. It does not scale past that, and it walks out the door when they leave.

alkemy runs that loop for you.

It scans your key accounts every week (news, competition, market moves) and filters everything into signals. You get a personalized feed you can act on. For any signal, you hit "Want to meet" and alkemy finds the right person to talk to at that company about that specific signal, then hands you a meeting prep doc. Every claim is cited to a source, and when it cannot find an answer it returns a blank instead of guessing.

How it is built, for the technical folks:

- A LangGraph Deep Agents harness underneath. One supervising agent per account that fans out to tools and subagents.

- Memory at two levels: the whole organization, and each target account.

- Memory is Markdown-first. Each account gets its own filesystem (index.md, brief.md, learnings.md, signals/, people/) as the source of truth. Postgres holds structured metadata, and pgvector is purely a retrieval index, not the source of truth.

Where I am now: building out the evals. If you have opinions on evaluating agent quality for research and retrieval tasks, I would genuinely like to hear them.

What I would love feedback on:

1. Does this actually help you build relationships and trust, in a world where AI does a lot of the grunt work?

2. How should signals be ranked?

3. How accurate do the briefs need to be before you trust them?

New accounts start with 2,000 free credits, no card required. Happy to get into the architecture or the data sources. If you want more room to play, I will top you off: nitesh@alkemyos.com