Building safety · Pre-beta
CoreThread →
Compliance for the Building Safety Act's golden thread: 16 regulation engines with 81 tests over 641 regulatory entries, an offline-first mobile app with voice capture, and hash-chained inspection records.
What we deliver
Whatever the problem, the method travels.
Raising a team's pace with agents, readying a platform for the moment everyone arrives at once, planning a rollout against the real system. Different industries, different challenges, the same approach: start from the business goal, measure, find what really limits the work, change it, and prove the change.
01 · Product and commercial
Business goals are the top-level requirements.
Scope creeps when nobody can say which business goal a piece of work serves. Every epic, story and test should trace back to one, and work that can't gets parked.
21
features switched on the day the product owner signed them off
What we did
- Product ideas ranked by leverage, business impact against engineering effort, and taken to the product owner and CTO as options with a recommendation.
- A decisions register for business calls, so product decisions sit with the product owner, not with an engineer or an agent filling a gap.
- Finished features held behind flags until the product owner signed them off: 45 were waiting by 26 September, and 21 went live the day they were approved.
- Commercial cases built on measured numbers, such as the payback on dedicated build hardware, worked out from real CI spend and the share of jobs the hardware actually took.
- A platform licence read against the delivery plan, so the places where the two conflicted were raised before the work started.
02 · Human delivery
Servant leadership: the team's time is the thing to protect.
Most teams don't lack effort. They lose time to meetings that read tickets aloud, plans nobody checked against the code, and work that waits in QA with no one chasing it. The leader's job is to clear that out of the way.
What we did
- Lead by example: the delivery loop the team is invited to try is the one we run on our own products every day, with the numbers published.
- Share what we know: a team knowledge base anyone can query, a fortnightly engineering leads forum for knowledge exchange and mentoring, and every method written up in public.
- Change by collaboration, not mandate: the team's delivery flow is voluntary, engineers keep what works for them, and their feedback reshapes it.
- Refinement moved before the meeting, and planning treated as a review: agents do the discovery legwork, and the engineer builds the plan and agrees the tests.
- Delivery measured in working hours, so overnight waits stop counting against the team.
03 · Agent delivery
An end-to-end automated loop.
Agents write code quickly. The hard part is trusting what reaches production, and running them all day without a person driving every step.
454
pull requests merged in twelve days on a live product, with no reverts
What we did
- A discovery to delivery loop: grounded briefs, tests written before the code, independent review, and releases that have to be earned.
- A team brain, repo maps and gotcha notes that every agent reads before it starts.
- Always-on builders on dedicated hardware, with one account per client so work never shares a login.
- A dispatch agent that raises work for a go, writes the briefs and watches the queue, but never builds, merges or releases.
04 · Performance under load
Find the real limit before your users do.
Systems rarely fail at average load. They fail in the minute everyone arrives at once, or in the one country where half the internet is blocked.
1,500
viewers simulated against production before show day
What we did
- Load-tested the production platform for a global corporate all-hands with 1,500 simulated viewers. The live caption service held every connection, with captions reaching 95% of viewers within 0.1 seconds.
- Found the real ceiling, about 60 page requests a second when everyone arrives in the same minute, and fixed it before show day: an early-opening lobby to spread arrivals, a cached welcome page that cut each arrival from five requests to three, and more CPU where it counted.
- Got the stream into China: tested from Beijing, Shanghai and Guangzhou, self-hosted the player that failed in Beijing, worked with distribution partners there, and raced two stream providers so each viewer gets whichever picture arrives first.
- Live captions in eight or more languages. The event ran without a fault.
- The same habit on our own hardware: measured at real concurrency, one DGX Spark delivers 120 times the throughput single-stream reviews report.
05 · Programme delivery
Plans that hold up against the real system.
A programme plan built from slides finds its gaps in production. One built from the code finds them on day one.
7
configuration gaps found by reading the deployment, not the plan
What we did
- Planned the rollout of a transport management platform to a new licensee. Diffing the two deployments' configuration found seven gaps before anyone built anything.
- Sequenced the work around the one thing nobody could speed up: app store review lead times for the rebranded driver app.
06 · Governance and due diligence
Knowing what you've agreed to, and proving you meet it.
Obligations live in contracts, questionnaires and email threads. The systems doing the work rarely know about them.
What we did
- An AI service review that mapped what each customer requires of the AI run on their content, such as approved models and processing regions, against what actually touches it, and flagged the conflicts.
- Supplier vetting that checks a supplier against regulatory, insurance and contract records, and flags discrepancies before work is placed.
- A vetting gate on every outgoing document: each line proven against current evidence, then checked again by an independent agent before it is sent.
07 · Where it has run
Six industries, one way of solving problems.
- Music rights and metadata
- Legal and claims
- Freight and logistics
- Live events and broadcast
- Building safety
- Skills and hiring
Three of those are our own products, built on the same loop. We say where each one is, because a site that claims more than it has shipped is the problem we fix.
Skills and hiring · Pre-launch beta
Leap →
Hiring on proof rather than CVs: translate experience into skills, prove them, then match. Three portals and signed skills passports the candidate owns.
Music industry · Live
PitchKit →
A WordPress plugin for trackable audio reels, so a music team knows who listened to what they sent.
08 · Start
Let's have a chat.
No pitch and no audit. Tell us what you're building and where you want to get to, and we'll think it through with you.