🚀 AI Engineering Portfolio Guide · 7 Projects · 7 GitHub Repositories

7 AI Engineering Projects That Actually Demonstrate Production Skills

If you want an AI/ML portfolio that goes beyond a chatbot demo, these projects cover the engineering layers hiring teams care about: retrieval, safety, model customization, reliability, observability, evaluation, and agent orchestration.

RAGText-to-SQLFine-tuningLLM GatewayEvaluationAgents
7 AI ProjectsOpen-Source References

7 AI Engineering Projects You Can Build

A practical project list covering retrieval, Text-to-SQL safety, fine-tuning, LLM infrastructure, prompt evaluation, RAG evaluation, and agent orchestration. Each idea includes a GitHub repository that actually demonstrates the core capability.

How to use this list

These are matching implementations, not necessarily the exact repositories that originally inspired the project titles. Use them as technical references. For your portfolio, build your own use case, modify the architecture, add evaluation and observability, and document the trade-offs.

RAG — RAG Pipeline with Hybrid Search →LLM Applications — Text-to-SQL with Guardrails →Fine-tuning — Fine-tuning Pipeline with LoRA →LLMOps — LLM Gateway with Rate Limiting + Fallback Routing →Evaluation — Prompt Versioning + A/B Testing Platform →RAG Evaluation — Automated Eval Dataset Generator →Agents — Agent Orchestration System →
1RAG

RAG Pipeline with Hybrid Search

matt-bentley/LLM-RAG-Architecture

GitHub ↗

What it demonstrates

Combines dense vector retrieval and BM25 keyword search, reranking, Qdrant, and multiple LLM providers.

What to build

Build a domain-specific RAG assistant: ingest documents, chunk and embed them, run hybrid retrieval, rerank candidates, generate grounded answers, and expose citations.

Make it portfolio-ready

Compare vector-only, BM25, and hybrid retrieval. Measure retrieval quality, answer faithfulness, latency, and cost. Add metadata filters, evaluation data, and failure-case analysis.

Skills demonstrated

Embeddings, BM25, vector databases, Qdrant, reranking, RAG evaluation, LLM APIs.

2LLM Applications

Text-to-SQL with Guardrails

Shamratha/Text-to-SQL

GitHub ↗

What it demonstrates

Includes SQL-AST guardrails, read-only execution, hallucination detection, validation, and confidence scoring.

What to build

Let users ask business questions in plain English. Generate SQL, parse and validate it, restrict access to approved schemas, execute safely, and explain the result.

Make it portfolio-ready

Block destructive SQL and unauthorized tables. Add confidence thresholds, error recovery, test cases, and query/result traces.

Skills demonstrated

Text-to-SQL, SQL parsing, AST validation, database security, hallucination detection, confidence scoring.

3Fine-tuning

Fine-tuning Pipeline with LoRA

ikaganacar1/Unsloth-Easy-Finetuning-Pipeline

GitHub ↗

What it demonstrates

Provides a configurable LoRA fine-tuning pipeline with 4-bit quantization, training, testing, and model serving.

What to build

Fine-tune a compact open model for a narrow task such as classification, extraction, support responses, or domain-specific instruction following.

Make it portfolio-ready

Create train/validation/test splits, benchmark the base model, track hyperparameters, compare LoRA settings, evaluate quality, and explain memory/performance trade-offs.

Skills demonstrated

LoRA, PEFT, quantization, supervised fine-tuning, GPU training, model evaluation, serving.

4LLMOps

LLM Gateway with Rate Limiting + Fallback Routing

vahid8/llm-gateway

GitHub ↗

What it demonstrates

OpenAI-compatible gateway supporting multiple providers, rate limiting, retries, fallback routing, cost/latency tracking, and a dashboard.

What to build

Put one API layer in front of multiple LLM providers. Route by model, budget, latency, or availability while keeping a consistent API contract.

Make it portfolio-ready

Add per-user quotas, exponential backoff, health checks, circuit breakers, structured logs, cost dashboards, and explicit fallback policies.

Skills demonstrated

API gateways, rate limiting, retries, routing, observability, latency/cost optimization, multi-provider LLM architecture.

5Evaluation

Prompt Versioning + A/B Testing Platform

iiizzzyyy/promptmetrics

GitHub ↗

What it demonstrates

Provides Git-backed prompt versioning, evaluation, metadata logging, and A/B testing between prompt versions.

What to build

Create a prompt registry where teams publish versions, attach test cases and metadata, run experiments, and compare outcomes before promotion.

Make it portfolio-ready

Track prompt version, model, dataset, latency, token usage, score, and human feedback. Define measurable promotion criteria and regression checks.

Skills demonstrated

Prompt management, experiment tracking, evaluation, A/B testing, metadata logging, LLMOps.

6RAG Evaluation

Automated Eval Dataset Generator

amazon-science/auto-rag-eval

GitHub ↗

What it demonstrates

Automatically generates task-specific evaluation exams from a knowledge corpus and uses them to evaluate RAG systems.

What to build

Build an evaluation pipeline that creates representative questions from your corpus, runs your RAG system, scores results, and stores regressions over time.

Make it portfolio-ready

Keep a reproducible evaluation set, separate generation from scoring, inspect false positives/negatives, and show before/after metrics when retrieval or prompting changes.

Skills demonstrated

Evaluation dataset generation, RAG testing, automated scoring, regression testing, experiment tracking.

7Agents

Agent Orchestration System

langchain-ai/langgraph

GitHub ↗

What it demonstrates

A strong foundation for stateful, long-running agents and multi-agent workflows with orchestration, memory, human-in-the-loop, and deployment capabilities.

What to build

Create an agent workflow with explicit states such as plan, act, inspect, retry, approval, and finalize. Add tools and persistent state instead of relying on one giant prompt.

Make it portfolio-ready

Build a concrete research, reporting, support-triage, or code-review workflow. Show state transitions, tool calls, retries, human approval, traces, and task-success metrics.

Skills demonstrated

LangGraph, state machines, agent workflows, tool calling, memory, human-in-the-loop, multi-agent orchestration.

How to turn one of these into a strong portfolio project

Choose a real problem

Pick a business workflow with a clear user, input, output, and success criterion. Avoid a generic chatbot.

Measure the system

Track quality and engineering metrics such as recall, faithfulness, SQL validity, task success, latency, cost, or prompt win rate.

Show production thinking

Include security, retries, failure handling, logging, evaluation, deployment, monitoring, and architecture trade-offs.