This page used to read like a shop for “business owners”: Midjourney, Jasper, Speechify, Asana. That is not who I write for, and I had not used those tools for the work on this site.
I write for students and engineers who want an ML or data job, or who have an agent demo and need it to survive. The list below is ten tools I actually use for that. If I add a paid link later, I will say so at the top. Right now these are just the tools.
These AI tools help me work faster, stay organized, and get more done in less time. If you run a business or want to be more productive, the right tools can make a big difference.
Who this is for
You. The person reading Nucleusbox. Not a founder looking for a slide deck generator.
If you run a shop and want marketing images, this is the wrong article. If you want to train a model, move data, or run an agent, stay.
1. Official ChatGPT and the OpenAI API
I use chatgpt.com when I want to think out loud: draft a mapping note, unstick a traceback, ask “why would this SQL pass a demo and fail a follow-up.” I use the API when NucleusIQ needs a model inside a run. Those are not the same bill. Plus is a chat subscription. The API is metered.
When I used it: every week, for writing and for agent examples that go through nucleusiq-openai.
The limit: the chat box will invent a confident paragraph. I already watched it write a “unique” data-engineering post that was generic. I still check the diff, the SQL, and the numbers.
Link: chatgpt.com. If you are deciding whether to pay, I rewrote that as what the free plan includes, and when I pay.
2. Python and Jupyter
This is the desk. Anaconda or a plain venv, then a notebook. pandas, NumPy, scikit-learn. I still send beginners here before I send them to an agent framework.
When I used it: the original ML series on this site, and the 2026 beginner tutorial where the same churn problem is solved with sklearn and then with NucleusIQ.
The limit: a notebook is not production. It is how you see the leak. Do not stay in it forever, and do not skip it.
Link: Jupyter. Start on this site with the beginner tutorial or the from-scratch roadmap.
3. Ollama (local models)
I run local models when I do not want a per-token invoice, or when the code should stay on the machine. I wrote a full setup for Claude Code against Ollama with Qwen Coder. “Free” there means no API bill. You still pay in RAM, VRAM, disk, and time. A 1.5B model on a laptop is not a 70B server.
When I used it: local coding help, and as a provider path NucleusIQ is built to speak (Ollama is on the About page next to OpenAI, Gemini, Anthropic, Groq).
The limit: long repo-wide tasks die on small context and slow decode. I shrink the task. I do not pretend a local 7B is a production cluster.
Link: ollama.com and my Ollama + Qwen Coder guide.
4. A GPU, and the stack I put on it
When the model will not fit in a laptop, I rent a cloud GPU. The question is T4 vs A10G vs A100, and then vLLM vs TensorRT-LLM vs Ollama. I published the numbers and the memory formula here: How to choose a GPU for LLMs. A10G is the one I keep pointing people at for most serving. T4 is the budget card. A100 is for fat models or high concurrency.
vLLM is what I use when I want fast serving and frequent model swaps. Ollama is what I use when I want something I can start on a desk. I do not run a 70B farm all day.
When I used it: any serving job that will not fit on the laptop. I rent a cloud GPU for it. [[FILL: cloud host]]
The limit: a card that is idle is expensive. Quantization that your engine does not support is a paper saving.
Link: the GPU guide. A reader already asked for the training sequel. That is the next hardware post, not this list.
5. Informatica, or a Python pipeline
For more than five years I worked in Informatica. That is still how I think about sources, mappings, and “is this extract complete.” On this site I also teach the idea in Python: extract, transform, load, then argue ETL vs ELT. I do not need a new SaaS every year to explain a late file.
When I used it: production data work at Informatica; teaching the same failure modes in ETL and ETL vs ELT.
The limit: a pretty mapping does not fix a key that is not unique. The tool will run. The dashboard will lie.
Link: Informatica if you are in that stack; otherwise start with the ETL posts above and a Python script you can read.
6. NucleusIQ
This is the harness I maintain. Direct, Standard, or Autonomous. Tools, plugins, memory, streaming, telemetry. I use it to teach “the demo worked” vs “the scorecard passed.” The text-to-SQL agent on Chinook is the public example: Claude Haiku, read-only SQL tools, pytest, an AgentResult you can grade without a hosted eval product.
When I used it: every agent post on this site, including the Gemini mixed-tool crash.
The limit: it will not save a bad eval. If you have no tests, you have a demo.
Links: What is NucleusIQ, GitHub, pip install nucleusiq nucleusiq-openai, and the text-to-SQL eval.
7. vLLM
I use vLLM when the model has left the laptop and I need a server, not a chat box. Paged attention, batching, and a model I can swap without compiling a TensorRT engine. That is the serving path in the GPU guide: vLLM for frequent swaps, TensorRT-LLM when the model stays put and I want the last tokens per second.
When I used it: writing How to choose a GPU for LLMs, including how tool calling behaves. One tool call comes back. I chain the next call in NucleusIQ.
The limit: it will not beat TensorRT-LLM on a frozen model, and it will not finish several tool calls inside one turn. The loop stays in the agent.
Link: vLLM docs and the GPU guide.
8. scikit-learn
I use scikit-learn when the job is still a table. Split, pipeline, fit, metric. pandas loads the file. sklearn does the model. The churn lab on this site does that before anyone wraps the same problem in an agent. If the split leaks, the score is a lie, and a later NucleusIQ tool will just wrap the lie.
When I used it: the 2026 beginner tutorial, the same churn problem in sklearn and then in NucleusIQ.
The limit: one accuracy number on a holdout is not a model I trust. I want the classification report, the confusion matrix, and ROC-AUC. sklearn will not grade an agent trajectory.
Link: scikit-learn and the beginner tutorial.
9. GitHub
This is where the code lives. I push NucleusIQ there, and I clone it when I want the text-to-SQL eval on a machine I control. A post without a repo is a screenshot. Readers get the same command I use: clone NucleusIQ and run the notebook under notebooks/agents/text_to_sql_eval.
When I used it: the framework repo, the issues, and GitHub Discussions. The About page points at the nucleusbox org.
The limit: a tidy README can sit there while the scorecard is failing. I still clone and run the tests. Stars are not a result.
Link: github.com/nucleusbox.
10. Claude
I use Claude when the agent needs a model that is not OpenAI and the run has to be graded. The public text-to-SQL agent on Chinook calls Claude Haiku through nucleusiq-anthropic: read-only SQL tools, pytest, an AgentResult. The June 2026 autonomous run is the one I published. ChatGPT stays the desk for writing. Claude is the model inside that eval.
When I used it: evaluating text-to-SQL agents, model claude-sonnet.
The limit: a Haiku scorecard does not make the next follow-up free. A cheap model still invents a join. I grade the trajectory.
Link: Claude and the text-to-SQL eval.
What I cut, on purpose
Jasper, Midjourney, Gamma, Speechify, Asana, Fireflies, Napkin, UIzard. I named them once. I did not ship ML, data, or an agent with them. They do not belong on this page.
NotebookLM I have not used as part of NucleusIQ. I will not add it until I have.
What to open next
If you are learning: Machine learning tutorial for beginners.
If you will serve a model: GPU guide.
If the agent already “works”: text-to-SQL evals.
That is the kit. Ten tools I will still be using next month.
