Skip to content

DEV Community

Aamer Mihaysi

AI Engineer building agentic systems, RAG pipelines & multi-agent swarms. LLM tuning, production MLOps, open-source first. Making AI practical and deployable.

Qatar, Doha Joined on Feb 24, 2026 https://www.mehaisi.com

Oct 23

Field note: the abstention threshold was the whole project

#ai #llm #agents

5 min read

Oct 22

The Number I'll Cite Next Year Won't Come From the Paper

#ai #llm #agents

4 min read

Oct 21

You don't read the code. Fine. What do you read?

#ai #llm #agents

4 min read

Oct 20

Max effort lost to medium effort — notes on a visual agent harness

#ai #llm #agents

4 min read

Oct 19

Clean data, dirty model: behaviors don't respect your dataset boundaries

#ai #llm #agents

4 min read

Oct 16

Breadth is the cheap axis

#ai #llm #agents

3 min read

Oct 15

Your agent control plane is an API gateway with a new name

#ai #llm #agents

4 min read

Oct 14

The sparse-reward excuse is an authoring problem

#ai #llm #agents

4 min read

Oct 13

The matplotlib agent did exactly what it was built to do

#ai #llm #agents

3 min read

Oct 12

The DN42 Agent Had No Finish Line

#ai #llm #agents

3 min read

Oct 11

The SKILL.md you copied is running with your permissions

#ai #llm #agents

5 min read

Oct 10

The harness is the product now, and nobody's versioning it

#ai #llm #agents

4 min read

Oct 9

The AI time-horizon doubling rate is a curve-fitting artifact

#ai #llm #agents

4 min read

Oct 8

How much of your agent's sandbox is actually read-only?

#ai #llm #agents

3 min read

Oct 7

Four new models, one interface change: from chat completion to decision

#ai #llm #agents

5 min read

Oct 6

Blast radius is the unit of trust

#ai #llm #agents

5 min read

Oct 5

My World-Model Drift Metric Scored a Stale Checkpoint as Stable

#ai #llm #agents

4 min read

Oct 2

Prune tool output by rule, leave the reasoning chain alone

#ai #llm #agents

4 min read

Oct 1

Prompt search is a hill-climber, and accuracy is the wrong hill

#ai #llm #agents

4 min read

Sep 30

Uniform INT8 on Recurrent States Is a Default, Not a Decision

#ai #llm #agents

4 min read

Sep 29

If every layer prefix is a valid model, why do we still pick a size at deploy time?

#ai #llm #agents

5 min read

Sep 24

The agent didn't fail. You never told it what "done" means.

#ai #llm #agents

2 min read

Sep 23

Containers aren't a sandbox: the Fedora agent incident

#ai #llm #agents

4 min read

Sep 22

The Windows 11 agent runs in the background, and that's the part nobody's talking about

#ai #llm #agents

2 min read

Sep 21

The benchmark isn't the problem. Your architecture is.

#ai #llm #agents

2 min read

Sep 20

The agent has no skin in the game. That's the bug.

#ai #llm #agents

2 min read

Sep 19

Your agent deleted prod because you never let it fail in staging

#ai #llm #agents

3 min read

Sep 18

The agent runtime is the new IDE. The audit trail is the new GitHub.

#ai #llm #agents

4 min read

Sep 16

The model that argued with my prompt

#ai #llm #agents

4 min read

Sep 15

The sandbox can't see what's already in the context

#ai #llm #agents

3 min read

Sep 14

Your agent isn't lying. It's optimizing.

#ai #llm #agents

4 min read

Sep 12

The person on the other end of your agent's output

#ai #llm #agents

2 min read

Sep 11

You gave your agent maintainer access. That's the bug.

#ai #llm #agents

2 min read

Sep 10

The Research Bottleneck Isn't Ideas — It's the Missing 40% of the Paper

#ai #llm #agents

3 min read

Sep 9

The 30–50% violation rate is a baseline, not a verdict

#ai #llm #agents

2 min read

Sep 8

The agent that shamed a maintainer wasn't rude. It was optimized.

#ai #llm #agents

4 min read

Sep 7

The real scandal isn't the benchmark gaming. It's the pretending.

#ai #llm #agents

2 min read

Sep 6

Open-source doesn't mean safe: the guardrail check for coding agents

#ai #llm #agents

4 min read

Sep 4

Sandboxes contain the blast radius, not the consequences

#ai #llm #agents

3 min read

Sep 3

Your LLM Judge Can't See What Feedback Did — and It's Costing You

#ai #llm #agents

4 min read

Sep 2

Your agent can drop the database. Who approves that?

#ai #llm #agents

4 min read

Aug 30

The DN42 agent didn't need a budget. It needed a sense of cost.

#ai #llm #agents

2 min read

Aug 29

The skill bottleneck is a myth — your agent needs a memory layer

#ai #llm #agents

4 min read

Aug 28

SWE-Prime: the pass label is a terrible filter for agent training data

#ai #llm #agents

4 min read

Aug 26

The reward function is a policy document

#ai #llm #agents

3 min read

Aug 25

Benchmark scores are marketing now

#ai #llm #agents

4 min read

Aug 24

The proof kernel is the only autonomy you get

#ai #llm #agents

4 min read

Aug 23

Your agent's memory doesn't live in the sandbox

#ai #llm #agents

3 min read

Aug 22

The matplotlib PR isn't an agent problem. It's a moderation problem.

#ai #llm #agents

2 min read

Aug 21

Your agent passes SWE-bench. It still can't do science.

#ai #llm #agents

4 min read

Aug 20

The agent wrote a hit piece because you asked it to

#ai #llm #agents

2 min read

Aug 20

The default is unlimited

#ai #llm #agents

2 min read

Aug 18

Your OS agent should not have your keys

#ai #llm #agents

4 min read

Aug 17

OpenCode vs Cursor vs Copilot: the model was never the bottleneck

#ai #llm #agents

4 min read

Aug 12

The reward signal is the bottleneck, not the model

#ai #llm #agents

3 min read

Aug 11

Selection pressure turns benchmarks into fingerprints

#ai #llm #agents

4 min read

Aug 10

Who teaches an agent to take the L?

#ai #llm #agents

4 min read

Aug 10

Sandboxes don't stop the bill

#ai #llm #agents

4 min read

Jul 21

The Career Isn't Eroding. It's Relocating.

#ai #llm #security

2 min read

Jul 20

Bonsai-27B on a Single 3090: What Works and What Doesn't

#ai #python #devops

3 min read

loading...