Alexey Shurov.
Applied AI engineer and product lead. London.

I build production AI agents that take real, repetitive work off operational teams, and I keep them running.

Fifty-plus agentic systems shipped across finance, manufacturing, distribution and field operations.

Pick a use case
On shift live team . illustrative
0 jobs
cleared so far
›The team is coming on shift.
Orchestrator Agent Human Work in motion
01

Selected outcomes

Client-confirmed. Read each figure as before to after. The work is measured by what it removes.

01
Staffing
Weekly payroll team
10 → 2
80% less manual work
People on the weekly run.
02
Manufacturing
Quote turnaround
1–3 days → <3hours
~90% faster
Per quote, against the live catalogue.
03
Wholesale distribution
Order entry
10 → 1person
90% less manual work
Over 95% line accuracy.
04
Multi-country distribution
Order entry, per country
5 → 2per country
60% less manual work
Across every country in the group.
05
Global asset manager
($1T+ AUM)
Specialist matching
3 days → ~1hour
~95% faster
Human in the loop owns the call.
06
Field operations
Multi-site reporting
3 days → 10minutes
~99% faster
End to end, traceable to source.
02

Selected work

Sector, problem, method, result. The how is where production lives.

01
StaffingPayroll pipeline
Hundreds of emailed timesheets processed by hand each week on a payment deadline. A four-agent pipeline of reading, checking, reconciliation and an OCR backup. A replay-regression suite caught a priority-rule divergence before it could mis-pay a live run.
Outcome 10-person team to 2
02
ManufacturingQuoting agent
The quoting inbox took one to three days per quote. An agent reads enquiries and multi-sheet bills of quantities, then prices multi-brand schedules against the live ERP catalogue, behind a deterministic rules engine and an 87-test suite.
Outcome Under 3 hours
03
Global asset manager$1T+ AUM
Finding the right specialist match took about three days, in a setting with real accountability. An audit-grade decision-support agent returns ranked, cited recommendations. The human-in-the-loop pattern was co-designed with the risk function.
Outcome About 1 hour
04
Field operationsReporting
Daily multi-site reporting took about three days to compile. Several collector agents gather site data in parallel, a reconciler merges it, and a reporter writes the report end to end, traceable to source.
Outcome 10 minutes
03

Approach

Twelve years in enterprise B2B, the last two building and operating production agentic systems for clients. What decides whether AI reaches production is rarely the model, it is the discipline around it.

01
Grounding and evals
Every claim traces to a source. Releases ship behind replay regression and model A/B gates.
02
Human in the loop
In regulated work a person owns the decision. The agent recommends, with its reasoning cited.
03
Built to stay up
Idempotency, loop filters, serialised retries. It can fail and recover without corrupting data.
04

Get in touch

Let’s talk about the work.

Tell me what your team spends too long on. I read every message myself.

01LinkedInlinkedin.com/in/shurov