🧭 ML Job Compass

Get the machine learning job you want by learning the skills to get there.

Author: Christina Joslin

Are you a college student interested in getting involved in machine learning (ML) but don’t know where to start? Early in your ML career and want to prepare for mid-level roles? Or aiming at senior/executive-track ML positions? ML Job Compass helps you chart the path to your desired ML role. Using Natural Language Processing (NLP), Data Visualization, Large Language Models (LLMs), and Retrieval-Augmented Generation (RAG), the app lets you:

1. Explore required ML skills in current U.S. job postings (2025) by industry domain.

2. Choose a target domain, career level, and optional ML specialization to get a tailored skills checklist or phased roadmap.

---

Features

• Exploration (by domain): Top programming languages, frameworks/tools & cloud platforms, ML specializations.
• Personalization: Pick your target domain, level (Early/Mid/Senior), and optional specialization; get either:
  • Skills checklist (3–5 items with what/why/mini-project/resources), or
  • Phased roadmap (foundations → core projects → portfolio/interview, with named resources).
• RAG-backed guidance: Uses a local Chroma vector store of real job snippets to ground suggestions.

---

Prerequisites

Step 1: Install Ollama on your local machine.

Step 2: Pull an Ollama model for inference:

• GPU recommended: ollama pull granite3.3:8b (Granite-3.3-8B-Instruct).

Inference speed and instruction-following quality were evaluated across several models (including the latest GPT OSS 20B model). Empirically, Granite-3.3-8B-Instruct was a solid default for this application.

• CPU-only option: ollama pull granite3.3:2b (smaller model, slower but workable).

Step 3 (if using containers): Install Docker Desktop

Step 4: Create a .env file

• Copy sample.env into .env and set values for:

  MODEL_ID=granite3.3:8b
  OLLAMA_BASE_URL=http://localhost:11434

Hardware note: Aim for ~16 GB RAM and, ideally, a NVIDIA GeForce RTX GPU (author tested on RTX 4070).

⚠️Image size note: The base CUDA/PyTorch image used for RAG (i.e., top-k similarity search during inference) is large ( $\approx$ 15.35 GB). Ensure sufficient disk space.

---

Deployment Options

There are two ways to run the app, both relying on Ollama running locally:

• Option 1 (Docker Compose):
  Runs a single container hosting the Streamlit app. The app makes HTTP calls to the local Ollama server (OLLAMA_BASE_URL=http://localhost:11434) on your machine.

• Option 2 (Local Python):
  Runs the Streamlit app locally on your host machine (via streamlit run app.py). The app still makes HTTP calls to the local Ollama server.

In both cases, the Streamlit app and Ollama communicate over HTTP; the only difference is whether Streamlit runs in a container or directly on your machine.

---

Option 1 — Run via Docker Compose

❗Ensure Ollama is running locally, your model is pulled, and .env is set up before starting.

Build & start:

docker compose up --build

Rerun using the existing image

docker compose up 

Environment

• Local default: OLLAMA_BASE_URL=http://localhost:11434
• Set MODEL_ID to your chosen model (e.g., granite3.3:8b).

---

Option 2 - Run Locally

❗Ensure Ollama is running locally, your model is pulled, and .env is set up before starting.

Step 1: Create & activate a virtual environment

Step 2: Install dependencies

pip install -r requirements.txt

Step 3: Start the app

streamlit run app.py

---

Navigating the Dashboard

Exploration Tab:

Use sliders to view up to the top 10 programming languages, tools & cloud platforms, and ML specializations by domain (or across all domains). Interactive Plotly visuals show counts from current postings.

Personalization Tab:

Select:

• Industry domain
• Target career level (Early Career/Mid-Level/Senior)
• Optional ML specialization (e.g., Natural Language Processing (NLP), Machine Learning Operations (MLOps))
• Output style (Skills checklist vs. Roadmap with a preparation window (up to 24 months))

After making your desired selections, click Generate Output. The app uses the vector store (RAG) & your selections to produce actionable, domain-specific guidance.

❓Have any questions while making a selection? Check the help icon for additional information.

---

Troubleshooting

LLM errors/timeouts:

• Ensure OLLAMA_BASE_URL is reachable, your model is pulled (check using ollama list) and the Ollama server is running.

Docker image is too large/low disk:

• Prune unused images/containers and ensure >20 GB free space:

---

Directory

├── .streamlit_/
│   └── config.toml                      # Streamlit theme & UI settings 
├── chroma_store_/                       # Persistent Chroma DB artifacts (created by ingest)
├── data_prep_/                          # One-off data prep & evaluation assets
│   ├── parse_job_info.py                # LLM parser for structured fields from raw job text
│   ├── parsed_test_ml_jobs_us.csv       # Sample parsed output (test subset)
│   ├── sentiment_similarity_test_v_true.png
│   ├── test_ml_jobs_us.csv              # Small raw test subset (CSV)
│   ├── test_ml_jobs_us.xlsx             # Same test subset (Excel)
│   └── Test_Parsing_Job_Info.ipynb      # Notebook for initial parsing/evaluation
├── .dockerignore
├── .gitignore
├── 1000_ml_jobs_us.csv                  # Raw jobs dataset (~1k postings)
├── app.py                               # Streamlit app: ML Job Compass UI
├── chroma_store.py                      # Build/query Chroma vector store (ingest + helpers)
├── docker-compose.yml                   # Dev stack (app + Ollama, etc.)
├── Dockerfile                           # App container image
├── parsed_1000_ml_jobs_us.csv           # Cleaned dataset consumed by the app
├── sample.env                           # Example environment variables 
└── requirements.txt                     # Python dependencies

---

Data Source

The original dataset used in this app, “Machine Learning Job Postings in the US” by Ivan Kumeyko, is available on Kaggle:
https://www.kaggle.com/datasets/ivankmk/thousand-ml-jobs-in-usa

The dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).

---

Notes & Recommendations

• Model choice: granite3.3:8b is the default. Smaller models (e.g., granite3.1:2b) can run on CPU-only machines.
• Hardware: A modern NVIDIA GPU with ~16 GB VRAM is recommended for responsive generation.
• RAG pipeline: If you change the dataset, re-run

  python chroma_store.py