AI Safety & Operations
Running the programs and systems underpinning AI Safety organisations
The development of Transformative AI is probably the most important thing happening right now, and that the field desperately needs people who can make organisations run. That’s the niche I work in: operations and systems for AI Safety organisations.
As the first employee at Catalyze Impact, I co-ran their incubation program for new AI Safety research organisations: recruitment, selection, program design, and logistics. At ML4Good, I taught sessions on AI agents, adversarial attacks, and AI Safety strategy at a 10-day intensive bootcamp. More recently I’ve been building CRM architecture and data infrastructure for an AI safety organisation.
The thread connecting all of it: small, cash-strapped teams can’t afford operational drag. I design the databases, pipelines, and automations (increasingly run by AI agents) that let a lean ops team punch above its weight, and I document them well enough that they outlive me.
I’ve written about my first year working in AI Safety operations, and I maintain a resources page for people who want to understand the field or move into it.
Projects in this area
All projects →CRM & Data Infrastructure for an AI Safety Organisation
Designed and built the central CRM architecture, alumni database, change-detection automations, and donor-import pipelines for an AI safety organisation
Catalyze Impact
First employee at the incubator for new AI Safety organisations. Co-ran the incubation program end to end
ML4Good EU March 2025
Taught AI agents, adversarial attacks, and AI Safety strategy at a 10-day intensive bootcamp
An AI safety organisation needed its scattered contact, alumni, and donor data turned into one reliable system. I designed and built it.
CRM architecture and documentation. I designed the central base — People and Organisations tables linked many-to-many through join tables, with a stable primary key — then did a field-by-field documentation pass across the entire base: table guides, consistent naming conventions for consent-sensitive fields, and flagging dead fields for deletion. The goal was a CRM whose structure a new team member can understand without me in the room.
Alumni database consolidation. I merged two parallel fellowship-alumni tables into one, while deliberately keeping team-assessed hiring data in a separate 1:1-linked table — self-reported and team-assessed data have different provenance and different sharing rules, and the schema should enforce that distinction.
Change detection without a changelog API. Airtable can’t tell you what changed on a record, only that it changed. For a self-service alumni profile form, I built a snapshot-and-diff automation: each record carries a JSON snapshot of its fields, diffed on every form submission, writing one changelog row per submission split into “needs processing” and “other changes”. The team reviews actual diffs instead of re-reading whole records.
Donor import pipeline. An email-triggered upsert script handling ~1,500 donation records from a major effective-giving platform: CSV parsing, donor matching by email then platform ID, ID backfill, and create-or-update keyed on donation reference. The key architectural fix was moving a downstream webhook call out of the import script — which was blowing through Airtable’s 50-fetch ceiling — into a separate record-created automation that Airtable throttles naturally.
AI Safety needs more organisations, and new organisations fail for ordinary reasons: no co-founder, no funding network, no operational footing. Catalyze Impact incubates new AI Safety research organisations to fix that, and I was their first employee.
For a year and a half I co-ran the incubation program end to end:
- Recruitment — finding promising founders through personalised outreach and social media campaigns.
- Selection — building and running the pipeline: reviewing applications and test tasks, interviewing candidates.
- Program design and logistics — shaping the program, arranging speakers and venues, and organising the events and trips that turn a cohort of strangers into a founder community.
- Participant support — being the person founders could bring any problem to during the program.
Because the team was small, I also built the systems that kept it running: Airtable bases and Python scripts that automated the repetitive parts of recruitment and selection, so the humans could spend their time on judgement calls instead of data entry.
ML4Good runs 10-day intensive bootcamps that take people from “interested in AI Safety” to actually working on it. I was a teaching assistant at the March 2025 Europe camp where I taught the sessions on AI Agents, Adversarial Attacks, and AI Safety strategy, and coached participants through the rest of the curriculum.
Because the atmosphere is just as important to a bootcamp as the atmosphere, I also ran the social side including a welcome event, murder mystery, lightning talks, and a Scottish ceilidh.