Principal AI Engineer. I catch agentic AI failures before they cost millions, not after. Evals + agentic systems for regulated enterprise. https://xitter.cf/t.co/wNKtykyGgV
New York, NY
Joined March 2009
- Tweets112K
- Following4.4K
- Followers4.7K
- Likes19.5K
My good sis @KanikaTolver just confirmed what I’ve been telling some people over the last couple of years.
Mike Legemah retweeted
Your review is judged against a rubric.
How many lines of it can you back with evidence?
LogPact's gap analysis shows which competencies have wins logged and which are empty, while there's still time to fix it.
Pro, 14-day trial: logpact.com/?utm_source=x&ut…
Prove a single-agent workflow before you build a multi-agent one.
Every extra agent is another handoff to secure and another failure mode to audit.
If you need several: delegation never grants permissions, and protocol compliance is not trust.
Why do non-US recruiters always think it’s a good idea to call you before they email you a job description? Do they think people are just sitting around doing nothing?
Mike Legemah retweeted
Unpopular opinion: the brag doc is bad advice.
Blank docs need you to remember them. You don't.
Smaller works: 1 win, 1 sentence, 60 sec, every day.
30 days → 30 dated wins. Free challenge starts Oct 26: logpact.com/challenge?utm_so…
LLM judge: "summary is grounded." 5-line script: "the claim ID it cites doesn't exist."
Believe the script.
Deterministic checks first. Judge for the rest, validated against human labels. Evals are release gates, not QA.
An agent's identity is not the identity of the person who asked.
Agent runs on a broad service account, user has limited access, agent fetches what the user can't see. Confused deputy.
Propagate the user's context. Authorize server-side at the resource, every time.
Don't give an agent one autonomy level for every task.
L1 read-only L2 reversible prep, human commits L3 controlled execution, approval + checks L4 high-impact, qualified human decides
Bind approvals to exact parameters. Change one, re-approve.
The accuracy lol
The people you'll meet on every interview panel:
- The Hiring Manager: decided in the first 5 minutes, still asks 45 minutes of questions
- The Peer: quietly worried you'll be better than him
- The Skip-Level: joins 10 minutes late, asks "Why this company?" and drops for another call
- The Culture Interviewer: no rubric, no notes, somehow has veto power
- The Shadow: in training, says nothing, still gets a vote
- The One Who Actually Read Your Resume: rare. If you get one, they're usually the reason you get the offer.
Mike Legemah retweeted
Panel interviews work like a jury trial these days.
One holdout and you don't get the job.
You meet about 3-5 people, and each one is assigned a piece of you to grade. One looks at your technical skills, one looks at how you work with a team, and so on. They score you on a rubric and then get together to compare notes.
Companies love it because it feels fair. Everyone is grading on the same scale, and no single manager can just hire their buddy.
But what most people don't know that is that at a lot of these companies the "yes" has to be unanimous, so four people can love you and one person can still prevent you from moving forward.
Mike Legemah retweeted
Stages of the modern day interview process:
1. June: "We're moving fast on this one"
2. July: "The hiring manager is on vacation"
3. August: "Just one more round with the VP"
4. September: "We're aligning internally"
5. October: "We've decided to move forward with other candidates"
6. November: the job is reposted
Mike Legemah retweeted
Your access ends on your last day.
Slack threads, tickets, email praise: all stay behind.
Your record of what you did shouldn't. Keep it independent of any employer, with proof attached.
Free to start: logpact.com/?utm_source=x&ut…
A confidence score is not a permission check.
Confidence says how sure the model is. It says nothing about whether it is right, or whether this user is allowed to do the thing.
Correctness: evals. Permission: code. Never trade one for the other.
The model proposes. Trusted code disposes.
In regulated AI the LLM is never the authority on what it may do. Validate the schema, authorize server-side, require human approval for high impact, then run it with an audit record.
mleg.tech
