How to work with me

18th Aug 2026

Two emails arrive on the same morning. The first, subject “Quick question”, hopes I’m doing well and asks whether I’d have time to discuss observability. The second asks whether head-based sampling drops LLM spans at 10%. I answer the second from my phone before the coffee is done. The first gets a reply asking what the question actually is, and we lose two days to that round trip.

Nothing about my calendar or my willingness to help changes between those two replies. Only the message does.

If you read nothing else here

Put your real question, with enough context to answer it, in the first message. Everything below is detail on that.

How to reach me

Messages come from readers, collaborators, recruiters and people stuck on a problem. I like getting them.

Email. My address is in the footer of every page on this site, or via the contact section. I read every note and reply to the ones I can help with, usually within a few days. If a week passes with no reply, I’m overloaded, not offended. A polite nudge is welcome and it works.

Put everything in the first message

Make your first message complete enough that I can answer it without asking you anything.

Ask the question, don’t announce it. “Can I ask you something about observability?” Yes, you can. Ask it. An email that only requests permission buys a round trip, and email round trips run in days, not minutes. Greet me and ask in the same message. (On Slack or LinkedIn this shows up as the bare “hi”, which is its own well-documented problem.)
Lead with the goal, not the step you’re stuck on. “I’m trying to trace token usage per tenant; I got as far as X and it breaks at Y” gets a far better answer than “how do I change this config value?” Often the step you’re fighting isn’t the right step at all.
Show what you already tried. What you searched, what you read, what failed and how. Narrowing a problem down before you escalate it is most of the work, and that context lets me skip everything you have already ruled out.
Be specific. Exact error messages, versions, a link, a minimal example. Vague asks (“it doesn’t work”, “any thoughts?”) stall; specifics get answers.

Don't send this

To: SoumendraSubject: Quick question

Hi Soumendra, hope you're doing well! I had a question about LLM observability. Would you be open to connecting?

Two days and one round trip before the actual question arrives.

Send this

To: SoumendraSubject: Does OTel span sampling apply to LLM traces?

Hi Soumendra! Quick one: does head-based sampling drop LLM spans too, or are they exempt? We sample at 10% and I can't tell if our missing generations are a bug. No rush.

✓ Answerable in one reply, even while I'm asleep in IST.

If AI helped you write it

Fine, I use these tools too. But edit it down to your words and your actual situation before you send it. If you paste raw model output, you’re sending me something I could have generated myself in four seconds, and a long unedited AI message signals that you didn’t invest the effort you’re asking of me.

Don't send this

To: SoumendraSubject: Synergistic collaboration opportunity 🚀

I hope this email finds you well. I am reaching out because…

If you didn't read it before sending, I won't read it after.

Send this

To: SoumendraSubject: Is eval-driven routing the right direction?

Hi Soumendra, I drafted this with Claude, then cut it to what matters: I'm building an agent that routes support tickets, my evals disagree with production outcomes and I've tried recalibrating thresholds. Is eval-driven routing even the right direction here?

✓ Cut down to your real situation, so I can answer it.

Calls and meetings

I work async-first, in IST (UTC+5:30), and I think best in writing, so a well-written email usually gets you a better answer than a call would. If something genuinely needs a call, propose it in the email with the agenda and the outcome you want, and I’ll share times.

Don't send this

To: SoumendraSubject: 30 min next week?

Hi Soumendra, I'm hitting a wall instrumenting our RAG pipeline and would love to walk you through it live. Are you free Tuesday or Wednesday?

Three emails to find a slot, for something writing would settle.

Send this

To: SoumendraSubject: Missing generation spans in LangChain RAG tracing

Hi Soumendra, I'm adding tracing to a RAG pipeline (LangChain + GPT-4). Spans show up for retrieval but not for the generation step; I've tried X and Y from the docs.

Minimal repro: (link). Am I missing an instrumentation hook? No rush, whenever you've got ten minutes.

✓ Answered in one reply, no meeting needed.

Don’t misunderstand me

A few things about how I communicate that can read wrong:

Short replies are not annoyance. I write tersely to answer fast; there is no subtext.
Silence means overloaded, not ignored. Nudge me after a week.
I push back on ideas out loud. Questions and counter-arguments are me engaging with your idea, not rejecting it.

What I’ll politely decline

To save us both a round trip:

“Can I pick your brain for 30 minutes?” Send the specific question in writing instead; you’ll get a better answer, faster.
“Can you debug my project / do my assignment?” I can point you in a direction if you show me where you’re stuck, but I won’t do the work.
Link exchanges, guest-post pitches and product promotion. No reply is coming.
Recruiter outreach that hasn’t read my profile. If you include the role, the range and why me specifically, in writing, I’ll read it.

Where this comes from

None of this is original. I collected it over the years from people who wrote it down first:

No Hello, on greetings that leave the other person waiting. I liked it enough to build my own version for people who ping me about observability platform access.
Don’t Ask to Ask, on asking the question instead of asking permission to ask it.
How To Ask Questions The Smart Way, Eric S. Raymond’s 2001 document that most of this genre descends from.
Stack Overflow’s guide to asking a good question, where the minimal reproducible example comes from.
Writing the Perfect Question by Jon Skeet, on what a badly framed question costs the person answering it.
Don’t Paste the AI and AI;DR by Rick Manelius, on the failure mode that arrived with the models.

I don’t always live up to this page. If your experience of working with me doesn’t match it, tell me. That is exactly the kind of message I want.