blog
Notes on AI delivery, engineering and setting solutions up inside the tools teams already use.
writing
- I re-engineered retail bank complaint handling using an autonomous agent workflow. Here is the exact architecture and the API costs.Seven agents prepare retail bank complaints for human sign-off at about 16 cents of API cost each, cutting staff time from 84 to a modelled 26 minutes.
- Letter of credit checking against a five-day clock: five agents, two human sign-offs and a 31-cent API bill per presentationFive Claude agents prepare letter of credit presentations for a checker's decision at about 31 cents of API cost each, cutting modelled staff time from 95 to 25 minutes.
- Security questionnaires drafted by agents, signed off by people: the architecture, the review points and the API billDrafting 80 security questionnaires a month from an approved answer library with four Claude agents costs about $118 a month in API fees, modelled to save 600 hours.
- I re-engineered AML alert triage and investigation using an autonomous agent workflow. Here is the exact architecture and the API costs.A six-agent workflow that prepares AML alerts for analyst decision at about 6.5 cents of API cost each, cutting staff time from 60 to a modelled 17 minutes.
- I re-engineered chargeback dispute handling using an autonomous agent workflow. Here is the exact architecture and the API costs.Handling 4,000 chargebacks a month with deterministic evidence code and four Claude agents costs about $176 a month in API fees, modelled to save around 1,400 analyst hours.
- I re-engineered freight invoice audit using an autonomous agent workflow. Here is the exact architecture and the API costs.Auditing 25,000 carrier invoices a month with deterministic rating code and four Claude agents costs about $292 a month in API fees, modelled to save around 3,000 staff hours.
- I re-engineered P&C claims intake (first notice of loss) using an autonomous agent workflow. Here is the exact architecture and the API costs.A five-agent FNOL intake workflow for 9,000 motor and home claims a month, modelled at about $0.28 of API cost per claim and $2,517 a month.
- I re-engineered lease abstraction using an autonomous agent workflow. Here is the exact architecture and the API costs.A four-agent lease abstraction workflow for 120 leases a month costs about $81 a month in API calls and is modelled to save 380 staff hours.
- I re-engineered prior authorization using an autonomous agent workflow. Here is the exact architecture and the API costs.A multi-agent prior-authorization workflow for a specialty practice doing 3,000 requests a month. About $330 a month in API costs and a modelled 1,900 staff hours saved.
- I re-engineered SME credit memo preparation using an autonomous agent workflow. Here is the exact architecture and the API costs.A six-agent workflow that prepares SME credit memos for committee at about $1.20 in API cost per application, cutting staff time from 13.5 hours to a modelled 2.5.
- How an engagement worksFrom a 20-minute call to a scoped pilot, and how the solution ends up inside the tools your team already uses.
elsewhere
- Triaging 25 invoices with Clef is a 27B multimodal decision model from CloudflareI wanted to know whether a decision model would be useful to a finance process owner, so I built the smallest thing that could answer that: an accounts payable triage tool that reads an invoice, decidoct 3, 2026, substack, 1 min
- The bottleneck moved to the reviewers. You are not hiring them.Every AI risk on the list is individually survivable. The one that isn't is the loop they form — and it closes on a five-to-ten year clockoct 2, 2026, substack, 1 min
- Making Any Website Agent-Ready with WebMCPThe web was designed primarily for people.aug 21, 2026, substack, 1 min
- Production Multi-Agent Systems: Architecture Patterns for Enterprise AIA technical deep-dive into designing, orchestrating, and deploying multi-agent systems that actually work in production.aug 15, 2026, substack, 1 min
- AI Governance in Banking: The Playbook for Safe, Compliant, and Defensible AIBanks are not tech startups.aug 7, 2026, substack, 1 min
- LLM-as-a-Judge: The Complete Guide to Using AI to Evaluate AIYour model made 10,000 decisions yesterday. How many were right? “I don’t know” is not a production answer.jul 30, 2026, substack, 1 min
- The Real Cost of Running LLMs in Production: A Complete BreakdownYour CFO is about to ask "how much will this cost?" and you need a better answer than "it dependsjul 29, 2026, substack, 1 min
- Why Most Enterprise AI Projects Fail Before ProductionThe graveyard of enterprise AI is littered with brilliant proofs of concept that died on the vine.jul 29, 2026, substack, 1 min
- GenAI in Regulated Banking: What Actually WorksAfter shipping RAG + LangChain in production under PCI DSS, leading a 10-person GenAI team through external audits, and watching three separate “AI-first” initiatives stall on compliance rjul 26, 2026, substack, 1 min
- Building an AI Center of Excellence: A Strategic Blueprint for 2026Introduction: Why 2026 Is the Defining Year for AI CoEsjul 20, 2026, substack, 1 min
- Enterprise AI Playbook for BankingThe artificial intelligence landscape of 2026 would have seemed implausible just a few years ago.jul 15, 2026, substack, 1 min
- Building an AI Control Plane: Operating Thousands of AI Agents in ProductionMoving a single AI agent from a Jupyter notebook to production is an achievement.jul 3, 2026, substack, 1 min