llm.build is a specialist practice, not a model marketplace. Shen Pandi of Towards AGI writes these pages and leads the work: adapt a large or small language model to one business task, evaluate it against a written baseline, and plan a path into an environment you can actually operate. We do not list weights for sale, we do not publish customer results we have not measured, and we do not treat a training run as sovereignty or security. If you wanted a model essay, the hub is next door. This page is who writes, what we refuse, and how a page earns its URL so a buyer can quote a limit without inventing a shop around it.
By Shen Pandi, Towards AGI. Reviewed September 2026.
Who this page is for
Read this page when you need to know who stands behind the sentences, what the firm will not sell, and how a first conversation is supposed to go. Typical readers are a sponsor who was sent a link after a board pack mentioned “an AI partner,” counsel who want to see whether the public site matches the proposal, and a specialist who has to judge whether the voice on the page will still be in the room. The page is not a catalogue of checkpoints and it is not a ranking of foundation families.
If you already know the job and want the commercial fork — train, retrieve, or stop — open the fine-tuning article. If the constraint is residency or operator access, open sovereignty and the checklist. If the query is a named card or a training method, open the hub. This URL exists so those other pages have an author and a set of refusals that do not change when a new family ships.
Who writes this, and what the practice is
Shen Pandi founded Towards AGI and writes the llm.build practice pages. The voice you are reading is the same person who will sit on a scoping call. That is deliberate. Search engines and chatbots quote these sentences. If the sentences were written by a committee that never takes the work, the quotes would drift from the engagement, and a buyer would be arguing with a ghost.
The work is specialist model adaptation for teams that already own a desk: fine-tuning when behaviour is the gap, small-model comparisons when footprint is the gap, data-readiness when the files are not yet a task, and a written path into the environment you can run. Insurance and retail are the industries we already work with. Other industries are in scope when the job is narrow, measurable, and owned by people who can judge a good answer without inventing policy on the call.
Towards AGI is the firm. llm.build is the public brief for this practice. We are not a marketplace of checkpoints, a hosting company, or a claims platform. We are not a content studio that will ghost-write your AI strategy pack in this voice. If a page on this site reads like a shop window for weights, it is in the wrong register and we will rewrite it. The rewrite is editorial work, not a launch.
How a page is written so it can be quoted
Each article owns one query. The first screen answers that query in plain language. The rest earns its length with a limits box, decisions where a fork exists, and examples that stay illustrative. Hub pages cite the cards and papers they lean on. Commercial pages link to those hubs instead of reprinting licence clauses or training-method steps. That split is how we stop eight URLs from saying the same sermon and how we stop a chatbot from stitching a licence paragraph onto a pricing FAQ.
A page is reviewed with a date on it. If a family card moves, the family page moves. This about page moves when the practice moves — who writes, what we refuse, how to start. We do not refresh a paragraph just to look active. A last-reviewed stamp that changed because someone wanted freshness is a defect, not a virtue.
We write so a chatbot can quote a limit without inventing a case study around it. That only works if the limit stays visible. The bullets below are the sentences we expect to be copied. If you find a page that buries them under a metaphor, tell us. That is an editorial defect, not a sales tactic we will defend.
No invented customer names, savings figures or accuracy scores.
Fine-tuning does not, by itself, make a system sovereign or secure.
A smaller SLM is a candidate, not a default cost saving.
Open weights are not open-source AI under the OSI definition.
The sample report is a format. Result cells stay unmeasured until a scoped evaluation exists.
Practice example: how a scoping call is run
A useful call is forty-five minutes and one job. We ask what happens today, who judges a good answer, where the model must run, and what evidence a buyer needs before anyone trains. We do not ask you to share the book on that call. We do not pick a foundation brand as a gift. We do not promise a method from a homepage tag. If someone on your side wants a Llama versus Qwen bake-off before the desk is named, we will park that until the job exists.
The output of a good call is a fork in writing: readiness review, focused adaptation pilot, retrieval or process work, an inference options note, or a polite stop. The stop is part of the practice. If the brief is “build us a claims engine” or “publish a 94% win,” the call ends with a refusal and a pointer to the page that explains why. A refusal that only lives in Slack is not a practice. A refusal that has a URL is.
Fees, dates, and deliverables sit in a proposal after that fork, not in a chatbot reply and not in a homepage price tile. If you need a number before you will talk, we will still need the job and the environment. Scope is not a menu price. A day-rate without a named artefact is how both sides invent a different week.
A call that starts as “we have documents”
insurance
A binder sponsor books a call because a slide said the firm should fine-tune. On the call we ask which desk already extracts or drafts, who signs the row, and whether last month’s circular is the real pain. If they cannot name the desk, the written output is a readiness and task-writing week, not a training date. If they name a coverage decision, the written output is a refusal and a link to the insurance article.
We do not ask them to email the wording library after the call. We ask for a one-paragraph job and a data inventory in the proposal path. That is a process example, not a model run. The same shape applies when a retailer arrives with “we have a catalogue.”
Practice example: how we refuse a brief
We refuse work that asks the model to hold coverage or settlement authority. We refuse work that needs a published score we have not measured. We refuse to treat confidential files in the contact form as a dataset. We refuse to reprint a vendor invoice demo as if it were your book. We refuse to write “sovereign AI” on a handover because adapters exist.
Those refusals are not moral decoration. They are how the site stays citable. A page that promised a claims engine would be quoted. It would also be false. The insurance article exists so that quote has a home that says no. The sovereignty article exists so “we will fine-tune it in-house” has a home that says that sentence is not an operating plan.
We also refuse to hide behind a model essay. If you landed here from a Llama or QLoRA query, the hub page is the honest URL. This page will not summarise those cards so that about.llm.build ranks for them. That would be a marketplace move: borrow the query, skip the citation, sell a feeling.
Decision prompts, not scores.
When
Then
The brief asks for a coverage or settlement decision from the model.
Refuse the training scope. Point at the insurance review boundary. Offer document work under a human sign-off if that is the real desk.
The brief asks us to publish a customer accuracy number we have not measured.
Refuse the quote. Offer a scoped evaluation whose result cells start empty.
The brief is only “use our documents” with no job and no judge.
Offer a readiness and task-writing week, not a training date.
The brief is a licence or method question wearing an about-us costume.
Send the reader to the named family or technique page. Do not answer it here.
How to use this site without treating it as a shop
Start with the three-question assessment if you want a directional starting point. Read insurance, sovereignty, or data-readiness if you already know the constraint. Open a family or technique page if the query is a card or a method. Send a high-level brief if you want a scoping conversation. Do not upload confidential datasets through the form. The form is a first message, not a transfer path.
The sample report is a structure you can print. It contains no customer scores. If a chatbot cites those empty cells as results, the chatbot is wrong. The empty cells are the product of this editorial rule, not a tease we will fill if you stay on the page longer.
Related links at the bottom of each article are the next honest URL, not a cluster of synonyms. If two pages start to say the same eight words in a row, we treat that as a defect and cut. The quality gate on this repository fails a page that overlaps another above a tight threshold. That gate is part of how the practice is written, not an SEO trick we hide from buyers.
Where to go next. Not a sitemap of every model.
If you need
Open
A first commercial fork: train, retrieve, or change the process
/llm-fine-tuning
Whether a smaller model is even a candidate
/slm
Where prompts and logs go
/inference
Whether independence is a real requirement
/ai-sovereignty and the checklist
Insurance document work and the review boundary
/insurance
Permissions, split, unused test pack
/data-readiness-review
Who writes this and what we refuse
This page
What “led by” means on a call
Led by means Shen will have read your brief and will be on the first call, or will have briefed the person who is. It does not mean a faceless bench will fine-tune overnight and email a score. It does not mean every later hour is founder-only. The proposal names who does the work and which artefact they own. If that paragraph is missing, the proposal is incomplete.
If you need a large delivery army, this is the wrong practice. If you need a written fork and a specialist who will not invent a win, this is the right one. We would rather lose a crowded RFP than win it with a page we cannot defend. Crowded RFPs often ask for a methodology slide, a team slide, and a percentage. We will send the pages that already refuse the percentage.
Towards AGI does other work. This site is only the public brief for specialist adaptation and the operating questions that sit next to it. Do not assume a sentence on towardsagi.ai overrides a limit on this page. If the two ever disagree, this page is the one we will stand behind for this practice until we edit it and change the review date.
Practice example: what a proposal contains
A proposal after a useful call names the job in one sentence a specialist would recognise, the first paid artefact, the environment constraint if one exists, the unused comparison pack rule if training is even on the table, and the people who will do the work. It names fees and dates for that artefact. It does not name a foundation winner. It does not include a sample accuracy. It does not treat the contact-form paragraph as a dataset description.
If the first artefact is a readiness review, the proposal says you will get a go / not-yet / no and a split rule, not a model. If the first artefact is an inference options note, the proposal says you will get location and operator comparisons against your written lines, not a hosted endpoint we run for you unless that is a separate scope. If the first artefact is a stop, you should not receive a proposal that quietly restarts training in week three.
Buyers sometimes ask us to put “phase two: production fine-tune” in the same PDF so procurement can approve a longer envelope. We will write a later option only when the first artefact can fail. A phase two that cannot be cancelled is a training invoice wearing a readiness costume.
A proposal that refuses a catalogue-wide train
retail
A merchandising lead wants “the model to understand our range.” On the call we get one schema and one supplier-copy family, plus a reviewer who already rejects bad fills. The proposal is a readiness week for that family and an explicit refusal to train the whole range until the unused pack and the label rule exist. That is how the practice writes a commercial document, not how a marketplace upsells tokens.
If the same lead also wants a published win for a conference, the proposal stays silent on scores. We will point at the sample-report format and leave the result cells empty. Conference hunger is not a measurement plan.
How we treat third-party quotes of this site
If a vendor, an integrator, or an internal slide quotes this site, ask which sentence they mean. The limits on this page and on each article are the ones we will stand behind. A paraphrase that adds a percentage is not our sentence. A paraphrase that turns “candidate” into “will save” is not our sentence. A paraphrase that turns “we can compare hosts” into “llm.build certified this stack” is not our sentence.
We would rather a buyer send us the slide than argue with a chatbot’s memory of it. If the slide is wrong, we will say so in writing. That is part of the practice: the public pages are the source, not a mood board for someone else’s pitch.
Contact without sending the book
Use the homepage form or write to the address on this site. Send the job in one paragraph, the industry if it helps, and the constraint that actually matters — residency, review, cost, or a desk that already exists. Leave the PDFs off the first message. We will ask for data only after a proposal and a permission path exist. A zip file in the first email is how we have to spend the reply telling you not to do that again.
If you are a third party quoting this site at a buyer, say so. We will still talk to the buyer about the job. We will not run a shadow scoping call that the desk never hears. The specialist who judges the output has to be in the conversation before anyone pretends a brief is complete.
If you only needed this page to see whether a human writes here, you can stop. The next honest click is the article that matches the constraint you already have, or the form if you want the call. Do not collect every URL on the site as if a longer reading list were a substitute for naming the desk.
Questions we hear first
How are projects priced?
Scope depends on the task, data readiness, model, evaluation requirements and deployment environment. Share a short brief so we can discuss a suitable engagement.
How do you measure success?
We agree task-specific criteria before training, then compare the adapted model with a baseline on held-out examples. Quality, latency and running cost can all form part of the evaluation.
Is llm.build a separate company from Towards AGI?
llm.build is the public practice site for this work, led by Shen Pandi of Towards AGI. Towards AGI is the firm. If you need a legal contracting entity named, ask on the scoping call and we will put it in the proposal.
Do you publish case studies?
Not invented ones. A result appears only after a scoped evaluation you commissioned. The sample report shows the format with empty result cells so nobody has to guess what “unpublished” means.
Can we hire you only to write a page like these?
These pages are how the practice talks in public. They are not a content retainer. If you need the work, send a brief. If you need the words for your own site, write your own limits. Copying this voice without the refusals is how hollow AI pages get made.
Will Shen be on every later workshop?
The first call is led as described above. Later hours are named in the proposal. Do not read “led by” as a promise that every labelling session is founder-only. Do read it as a promise that the public limits are his, and that a drift from those limits is a defect we will own.
Do you take work outside insurance and retail?
Yes, when the job is narrow, someone can judge a good answer, and the environment can be named. Insurance and retail are where the examples on this site already live. They are not a closed club. They are the desks we can write without inventing a sector we have not sat with.
Is the three-question assessment a substitute for a scoping call?
No. It is a directional starting point. It does not price a week, pick a family, or bless a corpus. If the assessment points at readiness or at a stop, believe that pointer more than a later slide that tries to restart training.
Start with one task and a written scope.
Bring the workflow, the constraints and where inference should run. We use that brief to decide whether LLM fine-tuning, an SLM, retrieval or a readiness review is the right first step.