Hands working on a laptop in low light

← All work

A huge number of AI jobs, and every one has to finish.

I work on a platform where AI agents research, qualify and prepare accounts for B2B sales teams.

AI work that finishes

The product is not one clever prompt. It is long chains of AI jobs: find the company, read about it, check it against the customer's ideal profile, find the right person, prepare the message. Every step depends on a model that somebody else runs.

And AI is moving faster than the capacity behind it. On an ordinary day one provider times out, another answers that it is overloaded, a third does not offer the model in the region you need, and a fourth says you have reached your rate limit. None of that is a bug in your code. All of it is normal.

With a queue and a scheduled job, each of those failures is a lost job, a duplicate, or a person re-running things by hand. These chains run on a durable workflow engine, Temporal, instead. Every step's result is recorded as it completes. When a provider fails, the workflow does not die: it waits, retries with back-off or moves to another provider, then carries on from the exact step where it stopped, even after a deploy or a crashed worker. A long wait is simply a workflow that is still waiting.

The platform I work on routes each step across the main AI providers, with fallbacks. Work that can wait goes to cheaper capacity. Each kind of task runs on the cheapest model that passes its evaluation. Outputs are checked against a schema and settings are pinned, so the same input gives the same answer. My part is the durable workflows underneath and the compliance engineering, as part of the platform team.

one workflow · a bad day at the providerssample values
[01]step research the account: done, result recorded
[02]call first provider: timed out
[03]retry back-off, second attempt: overloaded
[04]fallback another provider answers; output passes the schema
[05]wait rate limit reached: the workflow sleeps until the window reopens
[06]resume worker restarted during a deploy: continues from step 05
[07]done every step finished once; nothing ran twice, nothing was lost

Durable execution turns provider failures into waits, retries and fallbacks.

Compliance and closed doors

I led the technical side of SOC 2 and ISO 27001 readiness: the controls engineering is responsible for, and the evidence that proves them. Evidence is collected where the work already happens: reviewed pull requests, release records, threat-detection findings, vulnerability fixes and machine-image rotation.

When the stream of new vulnerabilities outran manual triage, I built a cockpit over the compliance tasks: one AI-assisted session per task, evidence written to documents, export for the auditor.

The doors stay closed while all of this runs. Nobody reaches production directly: access goes through a gateway that issues short-lived certificates. Internal tools sit behind an identity check, off the public internet. Certification is the auditor's to grant; the groundwork is in place.

The same thinking is available for your organisation: compliance readiness and AI workflows in production.

Putting AI agents into production?

I am happy to talk through how I would approach it.

Get in touch