Referrals: the list was never the bottleneck

Another product that sells matching as referrals.

I tried Jack & Jill's automatic matching referral with my own scenario these past couple of days. A few days ago I wrote a post about on-device deployment, and someone in the comments asked who they should talk to about the pitfalls of model conversion. I didn't have an answer at the time, so I figured I'd throw this need at it. I put the conditions in, and the list that came out looked pretty legit — direction, public output, tags all lined up.

I clicked on the second one and immediately spotted the problem. It recommended someone to me on the grounds that they'd written articles about edge computing. I looked them up — they do hardware selection and thermal management, which is a completely different track from model quantization and operator adaptation. Keywords collided, but the actual work didn't.

It's not hard to guess the mechanism behind this kind of Agent — scrape public info for semantic matching, then rank by relevance. That step isn't hard engineering-wise; I could cobble together a prototype myself with the SDK, nothing more than retrieval plus ranking plus generating a reason. The hard part is the second half. The list is out — who's going to make the approach, and why would the other person reply to you? The value of a referral was never in finding the person, it's in that person being willing to engage. This layer the Agent can't replace — it doesn't know whether you and the other person share any common context, or whether you have something they want.

I just got hands-on with the cross-compilation toolchain 3 days ago, and my need is very specific — I just want someone willing to reply to me with three sentences. No matter how long the list is, if the response rate doesn't go up, it's worth zero.

My verdict is one sentence. Matching can be automated, referrals can't. The list is an entry point, not a product.

Right now it's stuck at the entry point.