Writing Resume Bullets for AI Projects

Projectsmedium25 min
Ask Cairnly about this lesson
On this page

What you will learn

  • Write resume bullets for AI/ML projects that show engineering judgment, not just model usage
  • Use the Action-Architecture-Outcome formula for every bullet
  • Transform vague "used X model" statements into compelling technical narratives

Why this matters

Every CS student in 2026 has an AI project on their resume. The difference between getting an interview and getting filtered is whether your bullets sound like "I followed a tutorial" or "I made engineering decisions." Hiring managers at AI-adjacent teams (and increasingly all teams) want to see that you understand tradeoffs — not just that you called an API.

The Action-Architecture-Outcome formula

Every AI project bullet should follow this structure:

[Action verb] + [what you built/designed] + [specific architecture/technical choice] + [measurable outcome or tradeoff]

Example transformation

Before (weak):

Used OpenAI API to build a chatbot for answering questions about documents.

After (strong):

Built a RAG pipeline using FAISS vector search and OpenAI embeddings, with a confidence-gated retrieval system that reduced hallucinated responses by 40% on a 500-document test corpus.

The "after" version shows you made decisions (FAISS, confidence gating) and measured results (40%, 500-document corpus).

Category 1: RAG and retrieval projects

Template:

Built a [domain]-specific retrieval system using [embedding model] and [vector store], achieving [retrieval metric] on [dataset size] documents with [latency] p95 response time.

Real bullet examples:

Built a campus FAQ retrieval system using sentence-transformers embeddings and ChromaDB, achieving 87% top-3 retrieval accuracy across 1,200 policy documents with 340ms average query latency.

Designed a multi-stage RAG pipeline with query rewriting and re-ranking that improved answer relevance by 25% over naive top-k retrieval, benchmarked on 200 manually labeled question-answer pairs.

Implemented hybrid search combining BM25 keyword matching with dense vector retrieval, reducing zero-result queries from 18% to 4% on a 3,000-document academic paper corpus.

Category 2: Classification and NLP projects

Template:

Trained/fine-tuned [model] for [task] on [dataset], achieving [metric] while [specific technical decision — data augmentation, class balancing, inference optimization].

Real bullet examples:

Fine-tuned a DistilBERT classifier for toxic comment detection, handling 6:1 class imbalance with focal loss and stratified sampling to achieve 0.82 F1 on the held-out test set.

Built a resume screening classifier using TF-IDF + logistic regression as a baseline, then compared against a fine-tuned BERT model — the simpler model achieved 91% of BERT's F1 at 50x lower inference cost, informing the production architecture choice.

Trained a sentiment analysis model on 15K Hinglish (Hindi-English) product reviews, using custom tokenization to handle code-switching, achieving 78% accuracy vs. 62% for the English-only baseline.

Category 3: LLM application projects

Template:

Designed [LLM-powered feature] with [prompt engineering technique / guardrail / evaluation method], reducing [failure mode] by [amount] across [test size].

Real bullet examples:

Built an AI code reviewer using Claude API with structured output parsing and few-shot prompting, reducing false-positive lint suggestions from 35% to 12% across a 400-PR test set.

Designed a multi-turn tutoring chatbot with conversation memory and topic-boundary detection, preventing off-topic drift in 94% of 50-turn test conversations using a lightweight classifier gate.

Implemented prompt injection detection for a customer service bot using a two-stage approach (regex pre-filter + classifier), blocking 98% of adversarial inputs in a 1,000-sample red-team evaluation.

Category 4: Computer vision projects

Template:

Built [CV system] using [model/approach] on [dataset], with [preprocessing/augmentation strategy] to handle [specific challenge], achieving [metric].

Real bullet examples:

Built a document layout parser using YOLOv8 fine-tuned on 2,000 annotated pages, with rotation and noise augmentation to handle real-world scan quality, achieving 0.89 mAP on the test set.

Designed a real-time hand gesture recognition system using MediaPipe for landmark extraction and a lightweight CNN classifier, running at 28 FPS on a laptop GPU with 95% accuracy across 10 gesture classes.

Words that signal engineering judgment

Use these to shift bullets from "I used a tool" to "I made decisions":

Weak verbStrong replacementWhy it's better
UsedDesigned, Implemented, IntegratedShows agency
CreatedBuilt, Engineered, DevelopedShows construction
AppliedEvaluated, Benchmarked, ComparedShows judgment
Worked onOwned, Led, ShippedShows ownership
ImprovedReduced [X] by [Y]%, Increased [X] by [Y]%Shows measurement

The "So what?" test

Read each bullet and ask: "Could any CS student who followed a tutorial write this?" If yes, add:

  • A specific number (latency, accuracy, dataset size, user count)
  • A technical decision you made (and briefly why)
  • A comparison (vs. baseline, vs. previous approach, vs. alternative)
markdown
## Before/After examples — copy and adapt these
 
### RAG project
BEFORE: Built a chatbot using LangChain and OpenAI API
AFTER:  Built a domain-specific QA system using LangChain with FAISS
        retrieval and OpenAI GPT-4, achieving 85% answer accuracy on
        a 200-question evaluation set with average 1.2s response time
 
### Classification project
BEFORE: Trained a model to classify images using CNN
AFTER:  Fine-tuned ResNet-50 on a 5,000-image medical imaging dataset
        with stratified k-fold cross-validation, achieving 0.91 AUC-ROC
        while maintaining <100ms inference time for deployment
 
### LLM project
BEFORE: Used prompt engineering to improve chatbot responses
AFTER:  Designed a chain-of-thought prompting strategy with structured
        output validation that reduced hallucinated citations from 23%
        to 6% across 300 test queries

Try it yourself

  1. Open your current resume (or start a blank document)
  2. Pick your strongest AI/ML project
  3. Write 3 bullets using the Action-Architecture-Outcome formula
  4. For each bullet, verify it passes the "So what?" test — does it contain a number, a decision, or a comparison?
  5. Read each bullet aloud — if it takes more than 8 seconds to say, it is too long. Split it into two bullets.

Common mistakes

  • Listing tools without decisions — "Used LangChain, OpenAI, FAISS" tells the reader nothing about what you actually built or why those tools were chosen
  • No numbers anywhere — even rough numbers (dataset size, latency, accuracy) are better than zero numbers. If you did not measure, go back and measure.
  • Claiming results you cannot defend — if your bullet says "95% accuracy," be ready to explain the test set, the metric, and why 95% is meaningful for that task
  • Writing 5+ bullets for one project — 2-3 strong bullets beat 5 mediocre ones. Each bullet should highlight a different aspect (architecture, performance, a specific challenge solved).
  • Burying the technical decision — lead with the decision, not the context. "Chose FAISS over Pinecone for zero-cost local development" is better than "After evaluating multiple vector stores, decided to use FAISS."

Quick quiz

Q: Which resume bullet best demonstrates engineering judgment for an AI project?

How to explain it in an interview

"Every AI project bullet on my resume follows a pattern: what I built, a specific technical decision I made and why, and a measurable outcome. For example, my RAG project bullet mentions that I chose FAISS over a managed vector DB for local development speed, implemented confidence-gated retrieval, and measured a 40% reduction in hallucinated responses. This shows I didn't just follow a tutorial — I evaluated tradeoffs."

Checklist

  • Rewrote all AI/ML project bullets using the Action-Architecture-Outcome formula
  • Every bullet contains at least one number (metric, dataset size, latency, or comparison)
  • Every bullet passes the "So what?" test — it cannot be written by someone who just followed a tutorial
  • No bullet exceeds 2 lines on a standard resume template
  • Read each bullet aloud and confirmed it takes under 8 seconds to say