Help the agent know what is important, what is not —
and prevent it from making silent decisions.

Isaac Kargar is a morning person, which is a small fact that turns out to explain a lot. He does the hard thinking early, when there is enough of him left to check the work. By evening, he says plainly, there is no energy for it.

He runs Nazmi, an AI consulting and implementation practice built on an unusual promise for the category: work with the person who builds it. No account layer, no slideware — the site's own headline is “Shipped, not slideware.” The work splits into two fields. Agentic systems and automation, where the pitch is to take judgment-heavy work a team currently does by hand and hand it to agents they can actually trust. And custom model training, where open models — Llama, Mistral, Qwen — are fine-tuned, distilled and aligned until they know a specific domain properly. In practice a client engagement can run anywhere across MLOps, model training, model deployment, agentic systems, RAG, knowledge graphs and memory systems, depending on what the customer actually needs. He describes it, without decoration, as consulting.

The credentials behind that are the kind that take a while to accumulate. Ten-plus years shipping production AI and ML. A PhD at Aalto in multi-agent reinforcement learning and decision-making — a field that has aged extremely well, given that the techniques for making several agents cooperate and choose are now being retrofitted into large language models by everyone at once. Principal AI Scientist at In Parallel. Co-founder and Chief AI Officer at Resoniks, where the product listened to machines for the sound of something going wrong, and where he co-founded a €2.65M seed round. Before that, Silo AI, on autonomous driving simulation. More than a hundred technical articles along the way.

Ask him why this, out of everything a person could do with a life, and the answer is disarmingly small. It satisfies two appetites at once — the need to build something, and the need to keep learning. That is the whole motivation. Everything else is downstream of it.

The learning appetite has been industrialised. Kargar runs a personal Telegram channel with around four hundred subscribers, fed by an agent he wrote that pulls in links from Twitter, YouTube, LinkedIn — papers, blog posts, any URL really — summarises each one into text and a short audio clip, and files everything in a local SQL database. He reads the summaries daily, picks the one or two that matter, and passes those on. He is deliberate about the sources: company accounts over personalities, Prime Intellect for reinforcement learning, and a stated resistance to whatever technique the timeline has decided to be excited about this month. It is a filtering machine, built by someone whose professional life is filtering machines, pointed at his own attention.

Which is the same problem he has been working on at a much larger scale.

That project is Lerim, open-sourced under the Apache licence and described in one line as a compiler that turns AI agent traces into reusable context. Not a transcript archive — a filter. When an agent finishes a job it leaves behind an enormous, mostly worthless record of everything it thought. Buried in there is the part with value: the decisions taken, the constraints discovered, the corrections a human had to make, the facts established. Lerim extracts those and keeps each one linked back to the session it came from, so the next agent starts with cited context instead of rediscovering the same things from scratch.

Building it taught him that the hard part is not the part everyone assumes.

Remembering is easy. Forgetting well is the engineering problem. What he landed on is a hierarchy that behaves uncomfortably like a mind: recent context held close and available, mid-term facts periodically merged with their near-duplicates, rarely-used items allowed to decay, and the long tail parked in a database that gets read only when something actually needs it. Some labs call the consolidation pass dreaming. The comparison holds better than it has any right to.

And then the finding that a certain kind of engineer will recognise immediately, because it is the opposite of what the field rewards you for announcing. He built the fully agentic version first. Then a smarter harness, one clever enough to prune its own context window. Then he threw both away.

Cheaper, steadier, and — the part that matters for what he does next — small enough to fit a model he can train himself. Because once a workflow is decomposed into defined steps, each step is narrow enough to distil: record what an expensive frontier model decides at that step, then fine-tune a small model to reproduce the behaviour for a fraction of the cost.

He has been publishing the receipts for that, step by step, in a series on the Nazmi blog. It starts unglamorously — a 0.8-billion-parameter model taught to operate inside a SQL tool-use loop, improving from one solved task out of 220 to forty-something. It learned the protocol; it was not yet competitive. Six parts later, a 9B open model has been post-trained from 78 solved tasks to 115 out of 220 — level with GPT-5.5 medium on the same benchmark, for roughly fifty euros of training. That is the argument he makes to clients about model size, made in public, with numbers attached.

It is also the argument he applies to himself. In another post, he handed Lerim's own memory-extraction pipeline to Claude Code and let it run twenty-four experiments across two rounds to optimise it. Quality went up 41%. Nearly half of that came from a single schema-level change — swapping one call type for another, one line of code — rather than from any of the prompt engineering. An agent optimising an agent, with the result written down either way.

But none of this is the thing he actually guards against. Underneath all of it sits one discipline, and it is not about models at all.

A silent decision is a choice nobody asked for and nobody saw. Tell an agent to go and train a model on a dataset and it will do exactly that. Somewhere in the middle it will make a hundred small determinations — what counts as valid, what gets dropped, what to do with the rows that look wrong — and none of them will be announced, because from the agent's point of view they weren't decisions. They were just the work. His example is specific and unromantic: he checked a dataset one morning and found roughly a thousand data points missing information the agent needed, which is why the training was quietly failing.

His fix is slower and less impressive than the demo: the agent does the step, but I verify the result before continuing. Break the job into pieces, look at what came back, then release the next piece. It costs him time. It also means the failure surfaces on the morning it happens rather than three stages downstream, wearing a plausible face.

He is honest about why this keeps needing to be re-established. Trust in a model drifts upward on its own — you use it, it works, you check less — and performance appears to drop precisely because you stopped looking. Media enthusiasm accelerates the drift. So he periodically resets his own expectations, deliberately, as maintenance.

The lesson generalises past agents, and he knows it.

There is no shortcut in which you skip the unscalable version. You do it by hand, badly, until you understand it well enough to name the steps. The naming is what makes delegation safe. Everything before that is hoping.

The same instinct governs when he builds at all. Research background, research reflex: look at what already exists before writing anything.

Sometimes what comes back is a startup idea he plays with and abandons. Sometimes it is a genuinely unfilled gap and a small open-source tool exists a week later, like the browser extension that reads text aloud, written because he was, in his own words, lazy to read all the text an agent handed him. He is unembarrassed about the motive. Most useful tools are built by someone avoiding a chore.

Given the last word, he spent it on the thing he says he has no answer for: the layoffs, and the developers watching this technology arrive and wondering what is left for them. He does not offer reassurance. He thinks companies will be smaller, and says so. But he is precise about what does not compress.

Anyone can issue the instruction. The judgment about whether the thing that came back is any good — that is the part still doing the work. Which is, in the end, the same claim as the silent decisions: the value isn't in producing the output. It is in being the one who checks.