Co-founder & CTO
Javier Leguina
Javier is a co-founder of flowscope. He was a founding engineer at ModelML (YC W24), where he built for the world's biggest banks, and before that an AI engineer at startups including Encord. He holds an MS in Machine Learning from UCL, where he researched hierarchical reinforcement learning. He writes flowscope's posts on agents, production reliability, security, and the mechanics of shipping AI into systems that were never built to be driven by software.
Posts by Javier Leguina
10 August 2026
Controls for autonomous actions: audit trails, reversibility, and human oversight
When an agent posts a journal entry or releases a payment, the question that matters is not average model accuracy but whether each action is logged, attributable, reversible, and overseen. Three named frameworks now say what to control, and they map cleanly onto engineering.
5 August 2026
The objection that the data is too messy, and why it argues for a different method, not against automation
Messy data is real and expensive, and it defeats rules that require clean input. That is exactly why the case for automation rests on a method that reads the data as it is and routes the doubtful cases to a person.
3 August 2026
Onboarding a customer or vendor without the manual file
Onboarding a counterparty is slow because the same documents get requested and rekeyed across systems, not because the verification rules are hard. Here is the redesigned workflow, with the decision kept human.
29 July 2026
The regulatory trajectory for enterprise AI agents
Compliance work has a lead time, so an operator deploying agents needs the direction and timing of the rules, not just today's text. Here is the trajectory, and why auditability designed in early is cheaper than auditability retrofitted.
24 July 2026
Three-way match, rebuilt: matching the order, the receipt, and the invoice before payment
The control that prevents overpayment is also the largest source of manual work in accounts payable. The redesign turns the matching step into the easy part and moves the real work upstream.
22 July 2026
The hallucination objection, answered
"Models hallucinate" is correct in the abstract and misleading as a blanket disqualifier. The rate depends on the model, the task, and whether the output is grounded in source text and verified, which is what turns it into a measured, managed quantity.
20 July 2026
Monitoring an agent after it ships: drift, regression, and evaluation in production
A deployed agent can change behavior with no change on your side, because the model provider updated something and did not version it. Here is the production machinery that catches it before your customers do.
13 July 2026
Working a receivables ledger down: a redesigned collections process
Late payment is a working-capital cost most operators carry without measuring it. We walk one redesigned collections workflow and the mechanism by which it closes the gap between what is invoiced and what is collected on time.
24 June 2026
What process redesign looks like inside an industrial staffing firm
A workflow-level walk through recruiting and onboarding at a staffing firm, where the delays in time-to-submit and time-to-fill come from, what an agent can run, and where a human stays in the loop.
20 June 2026
The employee-monitoring laws that decide how an observation agent can be deployed
A handful of facts about US law govern where a capture agent can run, and the bill most blogs called a 2026 law never passed. This is what's actually on the books, and why flowscope adopts the failed bill's principles anyway.
18 June 2026
The long tail of document variability is the whole job
"Automate data entry" names only the easy part of the work. Rules and template OCR leave a stubborn share of every document stack for humans to key, and the long tail of document variability is the whole job, where a language model plus a human earn their place.
16 June 2026
What the agent-reliability curve says about which workflows are automatable now
A mid-2026 reading of the METR task-length curve, turned into a workflow-selection rule for operators deciding what to automate this quarter and what to wait on.
14 June 2026
Writing back into a system that has no usable API
Most mid-market automation stalls at the integration surface, not the model. The engineering case for acting at the desktop a clerk already uses, walked through the accounts-payable queue and QuickBooks.
12 June 2026
Shadowing instead of surveillance, and why the difference is behavioral, not cosmetic
The objection to an observation agent is grounded in real evidence: monitoring backfires when its data disciplines individuals. Diagnostic shadowing inverts every variable the research blames.
10 June 2026
Rebuilding the month-end close, from trial balance to statements
The monthly close is the cleanest worked example in operations because the benchmarks are public. We walk one redesigned close end to end, from where the days actually go to what a finance team does with them back.
8 June 2026
What a capture agent records, and what it's built to throw away
The first thing operators ask about an observation agent is what it collects and where the data goes. We answer with the actual architecture: scoped capture, redaction at the endpoint and again before storage, minimum retention, and terms you can check at trust.flowscope.com.
6 June 2026
Why documented processes rot, and why watching the work beats reading the SOP
Standard operating procedures go stale because the real process lives in tacit knowledge that resists being written down. Observing the work recovers the layer documentation cannot.
4 June 2026
How production reliability gets engineered, and why a demo is not evidence of it
A pilot at eighty percent on a clean slice tells you almost nothing about whether the workflow runs unattended on Monday. The machinery that closes the gap, with named benchmark numbers.