Frontier models
The largest commercial models, used where their reasoning is worth the price.
All practices / 01
Arab navigators crossed oceans by reading the stars while the wind did the hard work. We build the same way. Engineers choose the destination, the route and the trade-offs. AI carries the load. Every change still passes the same reviews and the same pipeline.
Beyond the label
Everyone has an AI page now, and most of them describe the same thing: a chatbot. Real work starts with judgement. Which model fits this job, at which stage of the workflow, and where a plain rule, a database query or a person will do it better and cheaper.
The usual
How FR87 decides
Power without control
Before AI, a careless team could only move so fast, so its mistakes stayed small. AI changes the scale. Many companies are rushing out whatever the model produces, and that AI slop travels just as fast: code nobody understands, features nobody checked, products heading for a crash. The power is real, which is exactly why the skill to steer it matters more than ever.
A careless rider wobbles along, eyes on the phone, and rolls into a palm. Slow speed, a small bump, little harm done.
The same rider, now on a rocket. Enormous speed, no plan, no checks. It veers off course, and the crash is as big as the power.
The same rocket with a flight plan, checks and a crew who review every step. That is how we work: the same power, steered all the way to where you meant to go.
Why one prompt is never enough
Ask an AI the same question twice and you may get two different answers. It is not reading from a database. It predicts the most likely next word, again and again, which makes it brilliant and unpredictable at the same time. Like an archer, one loose arrow can land anywhere. Good aim comes from a clear target, practice and a steady hand.
1One prompt, no guidance
Arrows land all over the target. Sometimes impressive, never reliable.
Our own workshop
Our research and development team spends its days with the models themselves: running them, comparing them and pushing them until we know where each one shines and where it fails.
The largest commercial models, used where their reasoning is worth the price.
Base models adapted and tuned to a specific job and a specific vocabulary.
Models we can inspect, fine-tune and run anywhere, including fully offline.
Our own machines for running models in-house, so sensitive data never leaves the building.
Every model is scored on your real task before we recommend it, not on a leaderboard.
The first principle
In a majlis, good advice comes from everyone at the table sharing what they know. AI is no different. A single line gets a guess. A full briefing gets work that belongs in your product.
1Mtokens held in one conversation
All the material a senior engineer would ask for before touching the code, gathered into one conversation the AI keeps in mind.
One task, two ways to ask
Nobody here talks business before the coffee is poured. The same courtesy works with AI. Here is one real task, an open-source dependency that was forked years ago and never merged back, asked two ways.
A passing word
Sort out the problem with our forked library and suggest a plan.
With no idea which library, why it hurts or what you value, the AI returns a generic checklist.
A proper briefing
Years ago this project forked an open-source library. The fork froze it in place, and a great deal of our own work now sits on top.
Work out which upstream release we forked from and list every change we made since. Explain what each one does and where it is used today. Leave nothing out.
After that we will compare options, and you are welcome to add your own. These are the ones I am weighing:
- Replace what it does with a small, separate service.
- Rebuild it as a clean project with no open-source base.
- Adopt the maintained successor and keep our changes as a thin layer on top.
Think like a principal architect. A library like this should never have been forked. At most it should have been patched in place with tests around the patch, so upstream can move on and the tests still protect our changes.
Push back if any of this sounds wrong.
Now the AI knows the problem, how we like to build, three real options and the standard we expect. One message, a thorough study and a clear recommendation.
The feedback loop
A traditional delivery cycle moves at the pace of a caravan: steady, and weeks between oases. An AI-assisted loop moves like a falcon. By the time the caravan has made one stop, the falcon has flown out and back many times, and every flight is seen.
Minutesfrom a piece of feedback to the next version on screen
Not a single request and a single answer, but a conversation: context in paragraphs, a screenshot of what looks off, a fix in seconds, again and again until it is right.
The method
We keep exploring apart from building, so the feedback arrives before any production code is written.
From the requirements alone, the AI builds a throwaway, clickable prototype. We put it online and everyone reacts to the same screen.
Feedback within minutesEngineers break the work into release phases by technical dependency, so it reaches production step by step, not in one leap.
Led by engineersNow the AI holds the full picture: documents, prototype, feedback, codebase and review history. Production code follows and is reviewed like any other change.
The usual pull requestsReview notes are context too
Like the threads of Al Sadu weaving, review notes are stronger together. On one pull request we handed all seven comments to the AI at once, and one refactor answered every one.
A single refactor
All seven, woven together.
Then it went back through the same review as every other change.
Before the guests arrive
A good host has everything ready before anyone walks in. Making a project ready for AI is a one-time effort that pays back on every feature that follows. If the AI cannot start, run and test the project by itself, the engineer becomes the bottleneck.
your-project/
MakefileThe whole stack comes up with a single command, so the AI can run what it builds.
AGENTS.mdArchitecture, commands and house rules, read at the start of every session.
tests/The AI writes, then checks itself. Seconds of feedback, not minutes.
PROGRESS.mdNotes on progress and strategy that carry over, so knowledge builds up instead of resetting.
A typical journey
An illustration of how a large build usually runs with us: four stages, every one through the same review process as any other work.
Screens built straight from your requirement documents and put online, so everyone can react to the same thing.
Stage 1The main workflows, end to end, running on real data.
Stage 2The harder capabilities, built on ground that is already proven.
Stage 3Hardware, access or a third-party vendor: the parts that do not depend on code.
Stage 4Often, the only thing left to wait for is outside the code.
Everything that can be built in code gets built. Same pull requests, same reviews, same pipeline.AI inside your product
The same engineers put models into client products. We pick the tool the problem actually needs, measure it against a simple baseline and know what every request costs.
Forecasting, segmentation, scoring and classification on your own data. Often well-engineered features and gradient-boosted trees, sometimes deep learning, always compared with a plain baseline.
Classifying, detecting and inspecting images, and reading documents: scans, forms and invoices turned into structured data a system can rely on.
Classification, entity extraction, sentiment and search across text. Small models built for one task where they are cheaper and more predictable than a general one.
Search over your own documents, summaries, drafting and assistants inside the products you already run. Built with test sets, guardrails and a known cost per request.
Ingestion, feature and training pipelines, deployment and monitoring. The quiet plumbing that decides whether a model still works in six months.
Where it pays back
A selection of the problems we take on.
Read, sort and extract from documents in volume, with a person reviewing the cases the model is unsure about.
Group customers by behaviour and purchase history so offers, pricing and outreach suit each group.
Respond to questions from manuals, policies, tickets and wikis, with a link back to the exact paragraph.
Gather and clean public data from many sources, then surface the patterns that matter to the business.
Draft replies, triage tickets and summarise long threads inside the tools your team already uses.
Demand, load and risk forecasts that planners can read, question and override.
Tell us what you would build if feedback took minutes instead of weeks. An engineer will write back.