AI Copilot — Personal Autonomous Assistant

A fully local, privacy-first AI assistant with three specialised engines: Job Hunter, Code Assistant, and Knowledge Engine. Runs entirely on your machine via Ollama — no cloud API, no subscription, no data leaving your device.

---

Project overview

Most AI assistants are general-purpose chat interfaces. This copilot is purpose-built for an engineer's actual workflow: finding relevant jobs in a specialised market, solving domain-specific coding problems, and getting well-researched answers on technical topics.

All three engines share a single local LLM (Llama 3 via Ollama), a unified config file, and a consistent CLI interface — so you interact with one tool, not three separate scripts.

Why local LLM: CV data, job search strategy, and code are sensitive. Running everything locally via Ollama means zero exposure to third-party APIs — your data stays on your machine.

---

Three engines, one interface

🔍 Job Hunter

Autonomously searches the live job market, scores opportunities against your CV using RAG, generates tailored cover letters as PDFs, and sends ranked alerts to your phone.

Step	What happens
Search	DuckDuckGo queries for your target roles — no API key
CV matching	Your CV is loaded into ChromaDB; relevant sections retrieved per job via semantic search
Scoring	Local LLM scores each opportunity 0–100 with specific reasoning
Cover letters	PDF generated per top match — tailored to company and role
Notification	Push alert with ranked summary sent to phone via ntfy

💻 Code Assistant

Solves coding problems, debugs errors, explains code, reviews quality, generates boilerplate, and converts between languages — all saved to file.

Command	What it does
solve	Generates a working solution from a natural language problem description
debug	Diagnoses the root cause and returns fixed code with explanation
explain	Plain-English explanation of what code does and how
review	Structured code review: correctness, performance, readability, security
generate	Production-ready boilerplate from a specification
convert	Translates code between languages (e.g. MATLAB → Python)

🧠 Knowledge Engine

Answers any question using live web search + local LLM reasoning. Grounded answers with source citations — not hallucinated responses.

Command	What it does
ask	Answer any question using web-retrieved context
dive	Comprehensive multi-source research summary on a topic
compare	Structured comparison table + verdict for two options
explain-concept	Technical explanation with examples and common misconceptions
fact	Fact-check any claim against current web evidence
sum	Summarise any long text into bullet points

---

Project structure

ai-copilot/
│
├── copilot.py                        # Main entry point — interactive CLI or --mode flag
│
├── src/
│   ├── agents/
│   │   ├── job_hunter.py             # Job search, CV RAG, scoring, cover letter PDF, ntfy
│   │   ├── code_assistant.py         # Solve, debug, explain, review, generate, convert
│   │   └── knowledge_engine.py       # Ask, deep-dive, compare, summarise, fact-check
│   │
│   └── utils/
│       ├── config_loader.py          # YAML config → typed dataclasses
│       ├── llm_client.py             # Ollama wrapper with retry, JSON, table parsing
│       └── notifier.py               # ntfy push notification utility
│
├── configs/
│   └── config.yaml                   # Your personal config (CV paths, ntfy topic, queries)
│
├── data/
│   ├── cv/                           # Place your CV PDFs here (gitignored)
│   └── vectorstore/                  # ChromaDB store (auto-created, gitignored)
│
├── outputs/
│   ├── cover_letters/                # Generated PDF cover letters
│   └── code_solutions/               # Saved code outputs with timestamps
│
├── notebooks/
│   └── copilot_demo.ipynb            # Interactive walkthrough of all three engines
│
├── requirements.txt
├── .gitignore
└── README.md

---

Setup

Step 1 — Install Ollama and pull models:

# Install Ollama: https://ollama.ai
ollama pull llama3
ollama pull nomic-embed-text

Step 2 — Clone and install:

git clone https://github.com/PRATdoppelEK/ai-copilot.git
cd ai-copilot
pip install -r requirements.txt

Step 3 — Configure:

# Edit configs/config.yaml:
# - Add your name, email, LinkedIn
# - Set paths to your CV PDFs
# - Set your job search queries
# - Set your ntfy topic (optional)

Step 4 — Add CV files:

# Place your CV PDFs at the paths set in config.yaml, e.g.:
data/cv/Lebenslauf.pdf
data/cv/Projekterfahrung.pdf

---

Usage

Interactive mode (recommended):

python copilot.py

Direct CLI flags:

python copilot.py --mode job                        # Full job search run
python copilot.py --mode ask --input "What is RAG?" # Quick question
python copilot.py --mode solve --input "Write a Python function to parse ISO 8601 dates" --lang python
python copilot.py --mode explain --input "attention mechanism in transformers"
python copilot.py --mode dive --input "solid-state batteries 2025"

Interactive session examples:

Copilot › jobs
Copilot › solve Write an LSTM training loop in PyTorch for time-series regression
Copilot › ask What are the best vector databases for production RAG in 2026?
Copilot › compare FAISS vs ChromaDB
Copilot › explain-concept Kalman filter
Copilot › fact GPT-4 can process images natively
Copilot › debug
  [paste broken code, end with ---]
  [paste error message]

---

Requirements

langchain-community>=0.4.0
langchain-ollama>=0.2.0
chromadb>=0.5.0
pypdf>=4.0.0
duckduckgo-search>=8.0.0
reportlab>=4.0.0
requests>=2.28.0
pyyaml>=6.0.0

---

Key concepts

Local-first architecture: Every LLM call goes to Ollama running on localhost. CV data, job search queries, and code snippets never leave your machine.

RAG for CV matching: Rather than including your full CV in every prompt, ChromaDB retrieves only the most relevant sections per job listing — keeping prompts focused and LLM responses precise.

Structured prompting: Each agent uses carefully designed prompt templates that enforce specific output formats (pipe-separated tables, named sections) so responses can be parsed reliably without fragile regex heuristics.

Persistent vector store: The CV vector store is built once and reloaded on subsequent runs — no re-embedding on every session start.

ntfy notifications: A free, open-source push notification service. No account, no app registration — subscribe to any topic string in the ntfy app and the copilot pushes job match summaries to your phone automatically.

---

Author

Prateek Gaur — Applied ML Engineer | Agentic AI | TU Berlin M.Sc. LinkedIn · prateekgaur@gmx.de

---

License

MIT License — free to use with attribution.