Operator and systems builder. I turn AI experiments into workflows teams can actually run. Background in ops, finance & sales leadership. Speaking at NODES 2026
San Diego, CA
Joined April 2015
- Tweets515
- Following82
- Followers168
- Likes282
Pinned Tweet
16 days to @neo4j #NODES2026. My session in one line: flatten a branched conversation into a graph and you invent history.
Don't Flatten the Tree. Nov 12, 11:30am PST. Free and online.
I read 471 of my Claude Code sessions. Cache reads dominated the tokens. Pricing the same week told a different story. I checked the rewrites from long breaks and revived subagents.
danmercede.com/guides/where-…
I kept 90,000 bytes, passed a 171-byte selected view, and recovered the exact original. I keep omissions and storage refusals visible. I'm checking bytes, not a model bill.
Full guide: youtube.com/watch?v=lfxVYGI3…
My session: neo4j.com/nodes/speakers/dan…
Register free: neo4j.com/nodes?utm_source=S…
A bug fix with a green test can prove nothing. If the test also passed before the fix, it never touched the bug.
failclosed's new three-way proof checks that. Weak test: NOT_PROVEN, exit 1. A test that hits the $50 edge fails before, passes after, fails on revert. PROVEN.
Open source, MIT. The three-way proof is on main, not in the 0.2.0 release yet.
Repo: github.com/OrionArchitekton/…
Guide: danmercede.com/guides/the-fa…
The demo repo is a toy: one shipping fee function, two fix branches.
I changed the same file in two worktrees, ran Orion Worktree Report, and checked what overlap, partial coverage and exit one actually mean.
Full guide: youtube.com/watch?v=jOZgj8JN…
When my bot's answer is late, the easy fix is to send the job again. That's how one task turns into two.
The new guide runs it for real on a clean Windows machine: the second send gets refused and the first answer comes back.
Watch the full guide: youtube.com/watch?v=HytRSllx…
Written walkthrough: danmercede.com/guides/claude…
A rule all three of my new Grok Bot templates share: the bot drafts, you decide.
The outreach one never sends without your yes. The guide runner stops at payment, phone codes and anything public. The daily scan only reads.
Copy one, swap in your own product or topic, and tell me what breaks.
Listicle Outreach Pitcher: x.ai/bot/F0j4YSAv1HQYpeeIf_A…
Guide Runner: x.ai/bot/YSpwc9lMpyipbslTw1x…
Daily Leverage Scan: x.ai/bot/Keu0lpSVwt68nLuaJiT…
I published three Grok Bot templates I use myself:
1. Listicle Outreach Pitcher: finds "best of" articles that don't list your product and drafts a pitch for each. You approve every send.
2. Guide Runner: runs a tutorial step by step in the real tool and writes down what the screens actually said.
3. Daily Leverage Scan: read-only, up to 3 ideas a day from X, HN and Reddit, with sources and a rough cost-vs-return check. It never acts on them.
Listicle Outreach Pitcher: x.ai/bot/F0j4YSAv1HQYpeeIf_A…
Guide Runner: x.ai/bot/YSpwc9lMpyipbslTw1x…
Daily Leverage Scan: x.ai/bot/Keu0lpSVwt68nLuaJiT…
Same JSON schema, three providers. Gemini passes. OpenAI strict mode fails it on 4 rules. Anthropic fails it on 2: no minimum, no format.
schemafit repair writes a valid copy for each, and both lint clean. Exit 1 to exit 0, before a single API call.
Open source, MIT. pip install schemafit
Repo: github.com/OrionArchitekton/…
Guide: danmercede.com/guides/why-ll…
Repairs can be lossy. Dropping minimum means the total >= 0 check moves into your own code.
I read all 471 of my Claude Code sessions from last week.
97% of my tokens were cache reads, so my cache looked healthy. But the 3% that were cache writes cost almost as much as all the reads combined, and 94% of the big rewrites came right after a cache expired while I was idle.
Where it went, with the script to measure your own: danmercede.com/guides/where-…
My camera moved. The world stayed stuck.
I make the films for Elsewhere Unfolds on my own GPU, and I check three kinds of motion separately: the camera, the environment and the action. The new guide goes from one strange place to a finished 30-second film.
Watch the full guide: youtube.com/watch?v=o_-x318-…
Written walkthrough: danmercede.com/guides/qualit…
