◇

Career Vision

I'm at a deliberate inflection point — transitioning from platform engineering into AI safety. Three paths are in view. Each has a different theory of change and a different ask of my background.

AI Safety Research

Transitioning into technical AI safety research with a focus on ML security, robustness, and the security of agentic AI systems. Likely pathway: DPhil at Oxford AI Security or an intensive fellowship programme. My statistics training, 10+ years of engineering, and hands-on agentic AI work give me a concrete foundation — and my DevSecOps background brings a security-first lens that is underrepresented in ML research.

DPhil / OxfordML SecurityRobustnessAlignment

AI Safety Engineering

Specialising in AI security, ML infrastructure security, or building safety evaluation infrastructure — at labs like Anthropic, DeepMind, or dedicated safety organisations. This path offers faster time-to-impact and fills a genuine talent gap: there are far fewer safety-focused engineers than researchers in the ecosystem. My agentic AI technical leadership translates directly.

ML InfrastructureSecurity EngineeringSafety EvalsAI Labs

AI Governance

Combining technical depth with policy and governance work — either as a researcher at GovAI-type organisations, through policy fellowships, or in grantmaking roles that require technical evaluation capacity. My linguistics background and international social enterprise experience are comparative advantages for cross-jurisdictional AI governance work.

PolicyGovAIStandardsInternational
▲

Current thinking leans toward Path 1 or 2 — staying deeply technical while pivoting to safety. The core open question: is the marginal safety contribution greater from an engineer who ships immediately, or from one who invests 2–4 years to become a researcher? If you have a perspective, I want to hear it.

■

Learning Path

Structured upskilling in AI engineering and safety — building from foundations to production systems.

Active

AI Engineering

6 weeks · 45+ hours

Structured to take you from foundations to production. Build real RAG pipelines, evaluation frameworks, and agentic systems.

1
LLM Fundamentals & RAG Foundations Build a working RAG pipeline that answers questions from real documentation
▾
  • How LLMs work under the hood: pre-training, tokenization, attention, post-training
  • The RAG paradigm: when to use it, how to scope your project
  • Data ingestion: handling different document formats for your corpus
  • Your first end-to-end pipeline: index, retrieve, generate, test
  • Framework: RAG Project Scoping Framework

You build

Interactive Q&A system on MCP documentation

QdrantHaystackFastEmbedGemini
2
Chunking & Embeddings Test 7 chunking strategies on your data and find the winner
▾
  • Why chunking is the most important decision in your RAG pipeline
  • 7 strategies compared: naive, sentence, recursive, semantic, hybrid content-aware
  • Embeddings deep dive: how text becomes vectors, FastEmbed vs Voyage
  • LLM-as-Judge evaluation with side-by-side comparison dashboards
  • Framework: Chunking Decision Framework

You build

Ranked chunking strategy backed by your own evaluation evidence

7 StrategiesVoyage AIStreamlitLLM-as-Judge
3
Advanced Retrieval Optimise retrieval accuracy from 70% to 90%+
▾
  • Vector DB internals: how HNSW and approximate nearest neighbor works
  • Hybrid retrieval: combining dense and sparse search with Reciprocal Rank Fusion
  • Reranking architectures: when and why cross-encoders beat bi-encoders
  • Search space narrowing: metadata filtering and LLM-based routing
  • Framework: Retrieval Strategy Selection Framework

You build

Evidence-based retrieval strategy with 4 techniques evaluated head-to-head

Hybrid SearchBM25Voyage RerankerLLM Routing
4
Mastering Evaluation Build your own evaluation system with golden datasets
▾
  • The evaluation challenge: why measuring RAG quality is harder than building it
  • Synthetic test generation, LLM-as-Judge, deterministic semantic metrics
  • Building golden datasets from scratch when no ground truth exists
  • Cross-validating 3 independent evaluation methods to find where your system breaks
  • Framework: RAG Evaluation Strategy Framework

You build

Golden dataset + multi-method evaluation framework

RAGASDeepEvalCustom JudgesTriangulation
5
Production Engineering Deploy a production chatbot with caching, memory, and observability
▾
  • Production architecture: the real tradeoffs between latency, cost, and accuracy
  • Semantic caching with Redis for sub-50ms response times on repeated queries
  • Conversation memory, query rewriting, and intent-based routing
  • Observability, user feedback loops, Docker deployment
  • Framework: Production RAG Architecture

You build

Deployed production chatbot serving real requests

FastAPIRedisStreamlitOpikDocker
6
Agentic AI & Security Build a self-correcting RAG agent with adaptive routing
▾
  • The intelligence spectrum: from single API calls to fully autonomous agents
  • Corrective RAG: grading retrieval quality and self-correcting when it fails
  • Adaptive agents: confidence-based tool selection across multiple sources
  • RAG security essentials: injection detection, retrieval validation, output sanitization
  • Framework: Intelligence Spectrum Framework

You build

Self-correcting CRAG system + adaptive multi-tool agent

CRAGHaystack AgentsTavilyGitHub MCP
◎

Your Mission

The goal is to create legible output — to practice the craft, get feedback, find collaborators, test your fit, and improve understanding. This is the operating framework for everything that follows.

◆

The Cheap Tests Ladder

Cheap tests require the least effort, time, or resources to reduce your uncertainty. Start very short (<1 hour each), progress to short (1–10 hours), then long (10–100 hours), then very long. Each rung gives a stronger signal that you're a good fit — without sunk-cost commitment.

Very short (<1h each)

  • Talk to people further along the path
  • Read abstracts, blog posts, newsletters
  • Watch YouTube videos on technical content
  • Run a GitHub repo, reproduce math proofs

Short (1–10h each)

  • Read research papers and agendas
  • Reproduce a toy version of a paper
  • Write a short post; estimate timelines
  • Attend an ML conference or workshop

Long (10–100h each)

  • Read a book; complete an online course
  • Replicate and extend a paper
  • Do an Apart Sprint hackathon
  • Form an inside view on timelines

Very long (<1000h each)

  • Internships and residencies
  • Research fellowships (MATS, AISC)
  • Masters programme or DPhil
  • Independent research project
■

Next Steps Framework

Read / Listen / Watch

Follow your nose through the resources below. Prioritise things that build intuition before depth.

Do Stuff — Create Legible Output

Summarise what you read. Write opinions. Code and do maths. Add to GitHub. Post to EA Forum or LessWrong to get feedback.

Network — Learn, Don't Sell

Reach out to people doing the work you want to do. Learn about their path, get feedback on your understanding, build relationships before you need them.

Apply — Even for Feedback

Every application is a cheap test. A rejection with feedback is worth the effort. Use jobs.80000hours.org and the EA Opportunities board.

▲

Roadmap

Concrete next steps toward AI safety — structured by timeframe and grounded in 80,000 Hours career advising.

Now

Immediate Actions

Near-term

1–3 Months

6 Months

Medium-term

◇

Opportunity Roles

Specific roles identified through 80,000 Hours advising as strong matches for this career transition.

Security Lead, Agentic Red Team

Anthropic

↗

Deputy Director, Research Unit — AI

Policy Institute (UK)

↗

Machine Learning Research Engineer / Scientist

UK AI Safety Institute

↗

Research Scientist, Science of Evaluation

DeepMind

↗

Senior Distributed ML Engineer

Redwood Research

↗

Policy Director

Institute for AI Policy and Strategy (IAPS)

↗

Governance & Policy Fellow

Centre for the Governance of AI (GovAI)

↗