GitHub lets you pin 6 repos. Pin 3 AI engineering projects. Everything else is secondary. Each pinned project should demonstrate a different layer of the stack.

What Makes a Project Portfolio-Ready

• Deployed and live — a URL you can click during an interview. Not "see the screenshots."
• Excellent README — architecture diagram, problem statement, demo GIF or screenshot, tech stack badges (see Tab 2)
• Quantified results — "RAG faithfulness: 0.91 on 30-question eval set" beats "good accuracy"
• Production code quality — type hints, tests, .env.example, no hardcoded API keys, no 500-line main.py
• Commit history — evidence of iteration, not a single "initial commit + everything"

The README is your project's cover letter. Most engineers write one paragraph. A great README takes 2 hours and generates 10x more recruiter interest.

# Project Name — One-Line Value Proposition

[![Deploy](badge-url)](live-demo-url) [![Python](badge-url)](pypi) [![License](badge-url)](license)

> **[Live Demo](https://your-demo-url.com)** | Built with Claude + FastAPI + ChromaDB

## What It Does
2–3 sentences. What problem does it solve? Who is it for? What's the result?
"A production-ready RAG API that lets you query your private documents in plain English,
with grounded answers and citations. Faithfulness: 0.91 on 30-question eval set."

## Demo
[GIF or screenshot of the app working — this is the most important element]

## Architecture
[Simple ASCII or Mermaid diagram showing: User → API → LLM/Vector DB → Response]

## Tech Stack
- **LLM**: Claude 3.5 Sonnet via Anthropic API
- **Retrieval**: ChromaDB + Cohere reranker (two-stage, retrieve-50 → rerank-5)
- **Backend**: FastAPI + Celery + Redis
- **Deployment**: Docker + Railway

## Results
| Metric | Score |
|--------|-------|
| RAG Faithfulness | 0.91 |
| Context Recall @5 | 0.84 |
| p95 Query Latency | 1.2s |

## Quick Start
```bash
git clone https://github.com/you/project
cp .env.example .env  # add your API keys
docker compose up
# Open http://localhost:8000/docs
```

## Key Technical Decisions
- Used two-stage retrieval (vector → reranker) because baseline hit_rate@5 was 0.65; improved to 0.84
- Chose ChromaDB over Pinecone for self-hosting cost control at this scale
- Added Anthropic prompt caching for system prompt (90% cost reduction on repeated queries)

# Generate a demo GIF with terminalizer or vhs
pip install terminalizer
terminalizer record demo
terminalizer render demo -o demo.gif

# Or use asciinema for terminal demos
pip install asciinema
asciinema rec demo.cast
asciinema upload demo.cast  # get a shareable URL

# Add shields.io badges to README
# https://img.shields.io/badge/Python-3.12-blue
# https://img.shields.io/badge/FastAPI-0.109-green
# https://img.shields.io/badge/Live_Demo-Click_Here-orange

# Pin your best repos on GitHub profile:
# Profile → Customize profile → Pin repositories → select 3

Structure (1 page, no exceptions under 10 years experience)

• Name + contact — email, GitHub URL, LinkedIn URL, portfolio/demo URL
• Summary (2 sentences) — role target + top 3 skills + years experience
• Skills — grouped: LLM APIs (Anthropic, OpenAI) | Frameworks (FastAPI, LangGraph, LlamaIndex) | Infrastructure (Docker, Redis, Prometheus) | Languages (Python, TypeScript)
• Projects — 3 bullet points each. Lead with impact, not description.
• Experience — reverse chronological. Reframe existing work using AI engineering language.

The STAR-Metric Formula for Each Bullet

# BAD — describes what you did
"Built a RAG system using ChromaDB and the Anthropic API"

# GOOD — leads with impact, shows measurement
"Built production RAG API serving 500+ queries/day: 0.91 faithfulness score,
1.2s p95 latency, 90% cost reduction via Anthropic prompt caching"

# BAD — vague agent project
"Developed an AI agent using LangGraph"

# GOOD — quantified with production details
"Shipped LangGraph research agent with circuit breaker (capped at $0.50/session),
human-in-the-loop approval for sensitive tools, structured logging — reduced
manual research time by 60% for 3 team members"

# BAD — generic Python experience
"Used Python for backend development"

# GOOD — specific AI engineering context
"Led migration of 4 microservices to async FastAPI with LLM integration;
added Prometheus metrics and Grafana dashboards reducing MTTD from 2h to 8min"

Keywords That Get Past ATS Filters

Include these naturally in your bullets and skills section:

# AI/LLM keywords (pick the ones you actually know)
RAG · Retrieval-Augmented Generation · LLM · Large Language Models
Anthropic Claude · OpenAI GPT · prompt engineering · fine-tuning
LangChain · LangGraph · LlamaIndex · vector database · embeddings
ChromaDB · Pinecone · Qdrant · pgvector · semantic search · reranking

# Infrastructure keywords
FastAPI · async Python · Docker · Docker Compose · Celery · Redis
Prometheus · Grafana · structlog · distributed tracing · OpenTelemetry
GitHub Actions · CI/CD · blue-green deployment

# ML/Eval keywords (if Track B)
HuggingFace · PEFT · LoRA · QLoRA · Unsloth · vLLM · GGUF
SHAP · model evaluation · evals · DeepEval · Ragas

Post 2–3 times per week during your job search. Share what you built, what you learned, what failed. AI engineering content gets 3–5x more engagement than generic tech content right now. One viral post generates more recruiter messages than 100 cold applications.

Content Formats That Work

• Build in public — "Built a production RAG system this week. Here's what surprised me about chunking..." [screenshot]
• Lessons learned — "3 things I got wrong in my first agent loop and how I fixed them"
• Quantified results — "Compared Cohere reranker vs no reranker on my test set. Faithfulness went from 0.65 to 0.91. Chart:"
• Technical takes — "Agents are overused. 70% of 'agent' use cases I've seen would be better as simple chains. Here's the decision framework I use..."

Warm Outreach Template (DM to AI engineer at target company)

Hi [Name],

I've been following [Company]'s work on [specific thing — blog post, open source project, paper].
I noticed [specific observation showing you did your research].

I'm an AI engineer with [X years] background in [your domain — networking, distributed systems].
I've been building RAG systems and LangGraph agents this year and published a few things:
[link to best project].

I'd love to hear how [Company] is approaching [specific technical problem they work on].
Happy to share what I've learned from my production RAG work if that's useful.

[Name]

Job Boards That Work for AI Engineering

• Anthropic Careers (careers.anthropic.com)
• AIJobs.net — AI-specific job board
• Levels.fyi — for compensation research before negotiating
• LinkedIn Jobs filtered by "AI Engineer" + "LLM" + location
• Twitter/X "hiring AI engineer" — many startups hire this way

System Design for AI — Common Questions

• "Design a document Q&A system for 10 million documents" → talk about chunking, vector DB selection, retrieval quality, caching, cost
• "Design an AI customer support system" → routing, RAG, escalation, human handoff, evaluation
• "How would you reduce LLM API costs by 50%?" → model routing, prompt caching, response caching, batching, smaller max_tokens
• "How do you evaluate a RAG system?" → faithfulness, context recall, hit rate, LLM-as-judge, eval datasets
• "How do you handle an agent that gets stuck in a loop?" → fingerprinting, max_turns, circuit breaker, structured recovery

# Framework for AI system design answers:

# 1. Clarify requirements
# "How many queries/day? What latency SLA? What's the document corpus size?
#  Is accuracy or speed more important? What's the budget?"

# 2. High-level architecture (draw boxes)
# User → API → [Retrieval: embed query → vector search → rerank] → [LLM: generate] → Response

# 3. Deep dive on the interesting parts
# "For retrieval, I'd use two-stage: ChromaDB top-50 then Cohere reranker to top-5.
#  Baseline hit_rate@5 is ~0.65; with reranking it's ~0.84 in my experience."

# 4. Scale and reliability
# "For 1M+ documents I'd use Qdrant or Pinecone. For cost, I'd cache
#  embeddings and responses. For reliability, retries with exponential backoff."

# 5. Evaluation
# "I'd run Ragas on a 50-question ground truth set. Monitor faithfulness drift in prod."

Coding Questions You'll Actually See

• Implement a retry decorator with exponential backoff and jitter
• Write a sliding window rate limiter
• Implement a simple embedding cache with LRU eviction
• Write async code: asyncio.gather for parallel API calls, return_exceptions=True
• Implement cosine similarity and explain why it's used for embeddings

Behavioural Questions with AI Framing

• "Tell me about a time you had to make a tradeoff between accuracy and cost in an AI system" — your cost monitoring and model routing work
• "How do you keep up with the rapidly changing AI landscape?" — your learning journal, the roadmap you built
• "Describe a production issue you've diagnosed" — your observability work, structured logging, Grafana alerts

Your grand capstone is a project that combines multiple Parts of this roadmap into one cohesive system specific to your professional domain. For you, that means networking / DPDK / telecom infrastructure — the domain where you have 4+ years of professional experience and can build something genuinely useful.

Suggested Capstone: Intelligent DPDK/VPP Documentation Assistant

A production AI system that lets network engineers query DPDK and VPP documentation, analyze configurations, and get code-level advice.

• RAG backend — index official DPDK/VPP docs, DPDK programmer's guide, API references, community mailing list archives
• Agent layer — research agent that can search docs, retrieve code examples, compare versions, generate configuration snippets
• Production API — FastAPI with auth, rate limiting, streaming, Prometheus metrics, Grafana dashboard
• CI/CD — GitHub Actions: lint → eval → Docker build → deploy on push to main
• Evaluation — 30-question eval set covering mempool, ring buffer, PMD, DPDK EAL, VPP node graph
• Cost monitoring — per-query cost tracking, prompt caching for large doc system prompt

Alternative Capstone Ideas

• Network configuration assistant: parse and explain complex DPDK/VPP configs with error detection
• Performance analysis pipeline: ingest benchmark results, generate reports, suggest optimisations via AI
• Telecom infrastructure knowledge base: RAG over 3GPP specs, ITU documents, and vendor documentation

Capstone Deliverables

• Public GitHub repo with excellent README, architecture diagram, and eval results
• Live demo URL (Railway, Render, or DigitalOcean — all have free/cheap tiers)
• Blog post or LinkedIn series: "How I built a production AI assistant for network engineers"
• HuggingFace Space demo (if Track B — fine-tuned model)