Work

AI Tooling for Audit Workflows

AI
Automation
Audit

I build and deploy specialized AI assistants that help auditors work faster and with less friction - automating the checkable core while leaving every judgment call with the professional.

A magnifying glass examining a checklist with verified ticks and one flagged exception, on a purple gradient.

What I do

I build and deploy AI assistants that help auditors work faster and with less friction. My role sits between the people doing the audit work and the AI technology underneath: I translate what auditors actually need into working tools, and I translate the technology back into something they can use without being engineers themselves.

The auditor always stays the decision-maker. The AI is support. No output ever claims a case is “approved” or “passes” review. The tools surface, structure, and flag. People judge.

How I think about it

The hard part of AI in knowledge work is verification. Code is easy for AI because it has an oracle: it compiles or it doesn’t, the test passes or it fails. Most professional work has no such instant signal, which is why fully autonomous AI struggles there and assistants, not agents, are the honest fit.

So I build on the parts that can be verified. Structured extraction you can check field by field. Spreadsheet logic that computes deterministically. Rule-based flagging where the rule is explicit. The mechanical, checkable core is automated hard; the judgment in the middle stays with the professional. That line is the whole design philosophy.

What I’ve built

A suite of specialized assistants, each scoped to one workflow rather than one giant do-everything bot. Each is grounded in the firm’s own reference material, runs in the firm’s language, and is tuned to avoid the failure modes that matter in a regulated setting: overconfidence, hallucinated sources, and quietly smuggling judgment into places it doesn’t belong.

Alongside the assistants, I build reusable automation recipes: take a messy export, reshape it into a standard layout, flag the exceptions, produce a clean reviewable artifact. Built once, used by everyone, the same way every time.

How I work

Direct and iterative. I test against real documents, not toy examples. I prefer brief, technically grounded analysis over polished summaries, and I’d rather ship a small honest tool than a big aspirational one. When a built-in feature already solves a problem, I use it instead of building a worse version.

The throughline: lower the barrier to good tools without lowering the bar for the work itself.