Resume Bullets for AI Projects: What to Write, What to Drop

Interview prepeasy12 min
Ask Cairnly about this lesson
On this page

What you will learn

  • Write resume bullets that pass the 6-second recruiter scan
  • Use the "Verb + What + How + Impact" formula for AI projects
  • Distinguish impressive metrics from meaningless ones
  • Avoid the most common AI resume mistakes

Why this matters

Recruiters spend 6-10 seconds on your resume. Your AI project bullets need to communicate "this person has built real things" instantly. A bad bullet like "Worked with LLMs and embeddings" says nothing. A good bullet like "Built a domain-specific RAG assistant achieving 85% precision@5 on 50 test queries, deployed to 200 daily users" lands you an interview. The difference is structure, specificity, and knowing what to drop.

The formula: Verb + What + How + Impact

Every resume bullet should follow this pattern:

[Strong verb] + [what you built] + [key technical details] + [measurable impact]

Example:

Built a domain-specific RAG assistant [what] using FAISS vector search and OpenAI embeddings [how], achieving 85% precision@5 and reducing customer support response time by 40% [impact]

Strong verbs for AI projects

Use verbs that show ownership and engineering depth:

Use theseNot these
Built, Designed, ImplementedWorked on, Helped with, Contributed to
Deployed, Shipped, LaunchedUsed, Learned, Studied
Optimized, Reduced, ImprovedFamiliar with, Experienced in
Evaluated, Measured, CalibratedResearched, Explored
Integrated, Automated, MigratedAssisted, Supported

Good bullets vs bad bullets

RAG pipeline

Bad: "Built a RAG pipeline using Python and OpenAI"

Good: "Built and deployed a 6-stage RAG pipeline with FAISS retrieval, confidence-gated generation, and safety guardrails, achieving 91% faithfulness on a 50-query evaluation set"

Why it's better: mentions the architecture, evaluation methodology, and a specific metric.

Bad: "Used embeddings for semantic search"

Good: "Designed a vector search system processing 100K+ document chunks with FAISS IndexFlatIP and bge-small embeddings, achieving sub-50ms retrieval latency with 85% precision@5"

Why it's better: includes scale, specific tools, and performance numbers.

Fine-tuning

Bad: "Fine-tuned a language model"

Good: "Fine-tuned Mistral-7B using LoRA (rank-16, 0.06% trainable parameters) on 500 domain-specific examples, improving task accuracy from 72% to 89% vs the prompted baseline"

Why it's better: shows the technique, efficiency, and a before/after comparison.

Evaluation

Bad: "Evaluated AI system performance"

Good: "Built an automated evaluation pipeline measuring precision@k, faithfulness, and answer relevancy using LLM-as-judge, running on 50+ labeled queries per deployment cycle"

Deployment

Bad: "Deployed an AI application"

Good: "Deployed a RAG assistant to Hugging Face Spaces serving 200+ daily queries with 1.8s p50 latency, semantic caching (30% hit rate), and automated quality monitoring"

What metrics to include

Include:

  • Precision@k, recall@k, MRR (retrieval quality)
  • Faithfulness score (generation quality)
  • Latency (p50 or p99)
  • Cost savings or efficiency gains
  • User count or query volume
  • Before/after comparisons

Drop:

  • "Accuracy" without specifying what it measures
  • Training loss (nobody cares about your loss curve)
  • Number of parameters (unless it's about efficiency, like LoRA)
  • Lines of code
  • Number of commits

What to drop entirely

Remove bullets that show consumption, not creation:

  • "Completed Coursera course on ML" — goes in Education, not Experience
  • "Familiar with PyTorch, TensorFlow, LangChain" — put in a Skills section, not bullets
  • "Read papers on transformer architecture" — unless you implemented something from the paper
  • "Experimented with different models" — experimentation without results isn't a bullet

Structuring your AI projects section

If your AI experience is a personal project (like the capstone), format it as:

text
RAG Assistant for [Domain]                          [Month Year]
Personal Project | github.com/you/project | live-demo-url.com

- Built a 6-stage RAG pipeline (safety → embed → retrieve → 
  confidence gate → generate → output check) using FAISS and 
  GPT-4o-mini, deployed to Hugging Face Spaces
- Implemented confidence-gated retrieval reducing hallucinated 
  responses, calibrated on 50 labeled test queries (precision@5: 
  0.85, faithfulness: 0.91)
- Added regex + LLM-judge safety guardrails handling PII 
  redaction and prompt injection detection

Three bullets is the sweet spot. One for the core system, one for the quality/evaluation story, one for an additional capability.

The "so what?" test

For every bullet, ask: "If the interviewer reads only this bullet, will they want to ask me about it?" If the answer is no, rewrite it.

  • "Used FAISS for search" — so what? Everyone has used FAISS.
  • "Calibrated a confidence gate achieving 90% precision with 12% refusal rate on production queries" — that's interesting. How did you calibrate it? What's an acceptable refusal rate? This starts a conversation.

Quick quiz

Q: Which resume bullet is strongest for an AI engineering role?

How to explain it in an interview

"I write resume bullets with the formula: strong verb, what I built, how I built it, and measurable impact. For AI projects, the key metrics are retrieval quality (precision@k), generation quality (faithfulness), and operational metrics (latency, cost, user count). I include before/after comparisons when possible. I drop anything that shows consumption rather than creation — courses completed, tools used, papers read. Every bullet should make an interviewer want to ask a follow-up question."