Offshorly Labs. The record of what our team builds and learns with AI.
Dated, numbered, written by the people who did the work, with the numbers before and after. Fixes on client work, releases of the playbook every team here works to, our fortnightly read on the AI model market, and plain guides for anyone deciding how to build with AI.
Every record dated, numbered and written the week the work was done.
How a record gets here
It happens on real work
A team fixes or builds something with AI on a client project or inside Offshorly. Before we offered this way of working to clients, we used it on ourselves.
It is written the same week, by the people who did it
What the problem was, what we tried, what we would not do again, and the number before and after. Where a client is involved, the numbers are the ones they have signed off, and the client is named only when they have agreed.
It goes here first
The short version goes to LinkedIn afterwards, and what we learned goes into the playbooks through the AI Excellence team, so it reaches every team, including yours.
The method behind the records, from training to delivery, is on one page.
See how we workSix records that show the range
Fixing AI Triage Without Rewriting the Checklist13 Aug 2026
Read the recordFixing Speaker Tagging by Rethinking What “Timestamp” Means7 Sep 2026
Read the recordChasing Two Map Bugs Below the JavaScript Bridge14 Aug 2026
Read the recordNew tricks in the playbook, v1.5 is published with skills and hooks.5 Jun 2026
Read the recordHow to Build an AI Development Team Without Hiring In-House16 Oct 2026
Read the recordThey Stopped Undercutting Claude. Now They Are Trying To Beat It.22 Oct 2026
Read the recordShelf by shelf, newest first
Five shelves, one running number. Quote a record by its number and date.
Technical Wins6 records
Fixes and builds on client work. Each one gives the setup, the constraints, the approach, the tradeoffs, the result and the lesson, in the words of the engineer who did it.
Fixing a Customer-Matching Order Problem in Our Document Pipeline
A customer reported one wrong number in one field. The real cause was a matching rule picking the wrong customer’s rule set entirely.
Fixing Speaker Tagging by Rethinking What “Timestamp” Means
A meeting assistant kept mixing up who said what. The fix was not a better comparison; it was comparing against a different kind of data.
Chasing Two Map Bugs Below the JavaScript Bridge
Custom pins broke in two unrelated ways at once, crashing iOS in 10 to 20 percent of sessions. Neither problem lived in our code.
Fixing AI Triage Without Rewriting the Checklist
One engineer built a command that replaces a guessed cause with a diagnosis. Daily tickets went from 140-plus to about 20.
How Structured Documentation Resolved Misalignment on a Complex Client Project
A team and a client learning the requirements together, and no single source of truth. What fixed the alignment.
Shipping features on a young codebase spread across many repos
An engineer inherited a hard feature build on a young codebase. The work was understanding it before writing anything.
Knowledge-Base2 records
How a piece of our own code works, explained by the engineer who owns it.
The retry loop that keeps a transcription pipeline honest
A small PHP trait that everything else in a meeting bot quietly depends on.
How we stopped iOS Maps from crashing
Map interfaces look simple until you render hundreds of points of interest under load.
Playbook1 record
A playbook is a tested way of doing a piece of work, kept in writing rather than in one person’s head. These are its releases. Every dedicated AI development team we build works to it, and we share it with our clients.
New tricks in the playbook, v1.5 is published with skills and hooks.
Nine skills that load themselves, five scripts that run on their own, and Playwright for QA.
AI Excellence12 records
Every two weeks our AI R&D team writes down what changed in the model market, what it costs in practice, and what we do about it in our own work.
They Stopped Undercutting Claude. Now They Are Trying To Beat It.
In August six releases measured themselves against Opus. Now two are trying to beat it, and what that changes for the work we actually do.
Nobody Tried To Beat Claude. They Undercut It Instead.
August produced no new frontier model, and a field that stopped competing on capability and started competing on price.
The Cache Got Cheaper. Cost Per Task Did Not.
Cache reads dropped 75 percent. Real for anyone billed by the token; for a seat holder, it never arrives.
Fable Became The Yardstick
July did not produce one AI story. It produced a comparison market.
Sonnet 5 Ships. Fable 5 Returns. Now What?
The cheaper workhorse, the returning flagship, and what each means for a development team.
The Rise And Fall Of Fable 5
Fable 5 lived for three days. The reasons it died are more interesting than the reasons it shipped.
When The Model Release Is The Risk
A new Opus six weeks after the last one. Why a fast release is a risk to plan for, not a gift.
Limits Are The Ecosystem Now
Weekly limits went up 50 percent. What that does and does not change for a team that works inside them.
Usage Limits Were The Point
A compute deal announced the day after we wrote that usage limits are the real bottleneck.
The Moat Is Shrinking. The Bigger Problem Is Usage Limits.
Which model is best matters less than it did a few months ago. What matters now.
The Vercel hack is a third-party AI warning.
AI is a multiplier for writing, research and software. It is also a multiplier for risk.
Claude was the best SWE model. Now people are asking questions.
For weeks it was clearly the model to beat for software work. That is why the mood around it stood out.
Guide2 records
Plain answers to the questions people ask before they build with AI.
How to Build an AI Development Team Without Hiring In-House
The gaps that stall an AI project, what each one costs you, and the three ways to fill them.
What is a forward deployed engineer, and should you hire one?
The title comes from Palantir. The way of working is now everywhere. What it means and when it is worth paying for.
We run Offshorly on tools built this way.
Before we offered this system to clients, we used it on ourselves. Four products came out of it, all built on the same playbooks.
Atlas
Our internal platform, built with the playbooks in three months.
It runs pipeline, delivery, finance and people from one live view, and agents act on what it finds.
Blueprint
From a brief to a proof of concept in days.
Reads a brief and its documents, raises the gaps and risks as interview questions, and produces the instructions for a working proof of concept.
Compass
Our training platform, built in a month.
Every engineer, designer and QA engineer comes through Compass before they join a client team, and the curriculum changes whenever the R&D team finds something better.
Zunou.ai
The agent harness our developers build.
We run our own operations on it and on Atlas, so our people spend their time on judgement and client work.
The playbooks, the training and the delivery system behind them are on one page.
How we workThe case studies are the whole engagement.
The records are the week-to-week view. Each case study gives the situation the client was in, what we built, what the team looked like, and what changed as a result.
EBI AI
A dedicated team for an AI product, four years and still growing.
Read the case studySpecialized Group
From an AI proof of concept to Moribian, a live product.
Read the case studyEvery engagement we can publish, in full: all case studies.
Questions about Labs
-
What is Offshorly Labs?1
The record of what our team builds and learns with AI: fixes on client work, releases of the playbook every team here works to, our fortnightly read on the AI model market, and plain guides. Each record is dated and numbered so it can be cited.
-
Is this your own work or your clients’?2
Both. Client work is described without naming the client unless they have agreed, and the numbers are the ones the client has signed off. Our own tools and the playbook are ours to describe in full.
-
How often is it updated?3
Records are added as the teams finish work, usually weekly. The AI Excellence newsletter comes out every two weeks.
-
Can I use these records?4
Yes. Quote them with the record number and the date, and link to the page. If you want this kind of work done inside your business, start with a free AI consultation.
Tell us the process that costs your team the most time.
A rough description is enough. Within one business day you will have our view on whether AI can take it on, what the first step would cost, and whether we are the right team for it.
- Honest assessment of whether we’re the right fit
- Fixed pricing, no hidden costs
- 10+ years of trusted delivery
No lengthy process.
Thank you for contacting us!