AI safety portfolio: build your first projects to get hired

Building an AI safety portfolio isn't just about showcasing technical skills. It's about demonstrating you understand the unique risks of advanced AI systems and can contribute to making them safer. As a machine learning engineer who's spent the last three years working on LLM infrastructure and safety guardrails, I've noticed a significant shift in what hiring managers look for in non-PhD candidates.

This guide shows you how to create portfolio projects that specifically address the skills needed for entry-level AI safety positions. It builds on earlier career path guidance by focusing on practical demonstrations rather than just qualifications.

1. Identify your target role within AI safety

AI safety isn't one job, it's a spectrum of positions with different skill requirements. Before building anything, decide which path suits your background:

• Technical implementation roles: Engineering-focused positions building safeguards into AI systems
• Evaluation & testing roles: Finding vulnerabilities and measuring alignment in existing systems
• Governance & policy roles: Developing frameworks and processes for responsible AI deployment
• Risk assessment roles: Identifying potential harms from advanced AI systems

Look at job descriptions across organisations like Anthropic, Conjecture, and ARC to spot patterns in what they request. Your portfolio should target these specific requirements rather than generic ML skills.

You'll know this step worked when you can name several specific AI safety organisations where your target role exists, and you can list their core technical requirements.

2. Choose projects that demonstrate AI safety-specific skills

Generic ML projects won't cut it here. Your portfolio needs to showcase understanding of alignment, interpretability, robustness, and safety-specific concerns.

Some effective project types for different roles:

For technical implementation roles

• Build a jailbreak detector for an open-source LLM
• Implement RLHF (Reinforcement Learning from Human Feedback) for a small model
• Create a tool to detect when models are reasoning poorly or hallucinating

For evaluation roles

• Develop red-teaming procedures for existing models
• Create evaluations that test for harmful outputs across various scenarios
• Build a benchmark that measures specific alignment properties

For governance roles

• Create an incident response plan for a hypothetical AI system failure
• Design a monitoring dashboard for tracking model behaviour in production
• Develop a documentation framework for model limitations and risks

Choose 2-3 projects that clearly align with your target role. Depth matters more than breadth here.

This step works when your projects directly map to skills listed in job descriptions, not just general machine learning capabilities.

3. Implement baseline techniques before innovations

I've seen too many portfolios that try to invent novel safety techniques without demonstrating understanding of established methods.

Start by implementing existing techniques:

• Constitutional AI approaches
• Red-teaming methodologies from papers by Anthropic or DeepMind
• Interpretability techniques like activation atlases or circuit analysis
• Robustness testing frameworks from published literature

Show you can understand, implement, and critically evaluate existing work before proposing your own innovations. Document your process thoroughly, highlighting both successes and failures.

You've done this step right when you can explain the limitations and tradeoffs of current approaches, not just their implementation details.

4. Document your projects with safety-focused framing

How you document your work matters as much as the work itself. For each project:

• Explicitly identify which AI safety concerns it addresses
• Discuss limitations and potential risks of your approach
• Include counterfactual analyses ("If this failed, what might happen?")
• Provide clear explanations of your methodology and reasoning
• Highlight ethical considerations throughout your process

Create repositories with thorough READMEs that foreground safety considerations rather than just technical implementation. Include your thought process and reflection on what you learned about safety.

This step works when a non-technical person could read your documentation and understand which risks your project addresses and why they matter.

5. Demonstrate engagement with the AI safety community

AI safety hiring managers want evidence you're engaged with the field beyond just technical work. Build this into your portfolio by:

• Contributing to open-source AI safety projects
• Participating in relevant AI safety contests and challenges
• Writing analyses of recent papers or developments
• Joining forums like the Alignment Forum or LessWrong (with thoughtful contributions)
• Attending workshops or conferences and documenting your learnings

Include these activities in your portfolio as they demonstrate commitment to the field and understanding of its collaborative nature.

Success here looks like having meaningful interactions or contributions you can point to, not just passive consumption of content.

6. Solicit and incorporate feedback

Safety work is inherently collaborative. Demonstrate this by actively seeking feedback on your projects and visibly incorporating it.

• Share your work in AI safety Discord servers
• Request feedback from practitioners on Twitter/X
• Present at local AI safety meetups
• Participate in peer review processes for community projects

Document the feedback you receive and how you incorporated it. This shows humility and willingness to improve, crucial traits for safety work.

You've done this right when your project documentation includes a "Feedback and Iterations" section showing how your thinking evolved.

7. Create a coherent narrative across your portfolio

The final step is tying everything together with a clear narrative that explains:

• Why you're interested in AI safety specifically
• How your background prepares you for this work
• What you've learned through your projects
• Which safety concerns most motivate your work
• Where you hope to contribute in the field

This narrative should appear in your portfolio landing page, project descriptions, and be ready for interviews.

A successful narrative helps hiring managers understand not just what you can do, but why you want to do this work and how you think about safety problems.

Place this content on a simple personal website and/or GitHub profile readme that links to your projects. Browse available job openings to ensure your narrative addresses current market needs.

When things go wrong

If you're not getting traction with your portfolio, these are the common issues:

• Too generic: Your projects look like standard ML work without safety-specific elements. Add explicit safety framing and techniques.
• Too shallow: You've implemented too many small projects rather than deeply exploring fewer concerns. Cut back and go deeper on 2-3 projects.
• Too theoretical: Your work is all analysis without practical implementations. Balance theory with working code.
• Too isolated: There's no evidence of community engagement or feedback. Start participating in communities even before your projects are finished.
• Misaligned with target roles: The skills you're demonstrating don't match what employers need. Revisit job descriptions and realign your focus.

The good news is that AI safety roles remain in high demand, particularly for those who can demonstrate practical skills rather than just theoretical knowledge. The Hubfluencer program can help you build visibility once you have portfolio pieces to share.

Remember that in AI safety, showing good judgment and careful thinking often matters more than technical brilliance. A thoughtful project that thoroughly addresses a narrow safety concern will generally outperform a technically impressive but safety-naive one.

And be patient, building a compelling portfolio takes time. Three months of focused work is a reasonable timeline for developing entry-level projects that can actually help you stand out in this competitive field.