Skip to content

HuckleR2003/hck-GPT

Repository files navigation

hck_GPT

v2.1.0 · Part of PC Workman HCK

AI diagnostic assistant for Windows system monitoring. Answers natural language questions about your PC — in Polish and English — using real hardware data. No cloud. No API key.


What it does

You ask. It checks your actual hardware. It answers.

"Why is my PC slow right now?"
"Is cs2.exe a virus or a normal process?"
"Which game pushes my GPU the hardest?"
"RAM na 91% - co powinienem zamknąć?"

82 built-in intents covering hardware diagnostics, performance analysis, process identity, driver status, gaming analytics, battery, startup programs, and system health.


Architecture

hck_gpt/
├── engine/
│   └── hybrid_engine.py     # Routes messages: rule engine first, Ollama LLM fallback
├── intents/
│   ├── parser.py            # Intent parser with Polish diacritic normalization
│   ├── vocabulary.py        # 82 intents, PL+EN patterns, confidence scoring
│   └── lang_detect.py       # Auto-detects Polish vs English per message
├── responses/
│   └── builder.py           # 5600+ lines of bilingual response handlers
│                            # Every handler uses real psutil/WMI/SQLite data
├── memory/
│   ├── session_memory.py    # Conversation context, CPU/RAM trend buffers
│   ├── proactive_monitor.py # Background daemon: CPU/RAM/GPU/disk/temp alerts
│   └── user_knowledge.py    # SQLite user profile (hardware, usage patterns)
├── context/
│   ├── system_context.py    # Live snapshot: processes, temps, averages
│   └── hardware_scanner.py  # WMI scan: CPU model, GPU, VRAM, mobo, RAM speed
├── data/
│   ├── live_sensors.py      # LibreHardwareMonitor bridge
│   └── metrics_store.py     # DeepMonitor 5-min snapshots
├── chat_handler.py          # Message routing + quick aliases
├── panel.py                 # Tkinter UI: Bordeaux Noir panel, TIP/HOT strips
└── insights.py              # Habit tracking, anomaly detection, teasers

How the hybrid engine works

Every message hits the intent parser first. Known intents (82 total) go to the rule engine — fast, predictable, deterministic, no GPU needed.

Low-confidence or open-ended messages get routed to Ollama (local LLM). The engine injects a 6-section system context into the prompt: live CPU/RAM/GPU, today's averages, top processes, temperatures, hardware profile, and conversation history.

Ollama unavailable? 60-second cooldown, graceful fallback. No crashes.

User message
     │
     ▼
Intent parser  (confidence 0.0–1.0)
     │
     ├─ >= 0.60 ──► Rule engine handler ──► bilingual response + live hardware data
     │
     └─ < 0.60  ──► Ollama LLM (local, port 11434)
                     + 6-section system context injected
                     │
                     └─ unavailable ──► structured fallback

Intent categories

Category Example questions Count
Hardware info "What GPU do I have?", "How much RAM?" 6
Diagnostics "Is my PC healthy?", "Check temperatures" 4
Performance "Why is it slow?", "Compare today vs yesterday" 6
Process identity "Is svchost.exe a virus?" 3
Gaming "Can I run Cyberpunk?", "Which game stresses hardware most?" 5
Startup / drivers "What starts with Windows?", "Are drivers updated?" 4
Resource analysis "Why is RAM high?", "What's the top memory hog?" 6
Time-travel "What changed since last week?", "Crash context?" 7
Battery / power "How fast is battery draining?" 4
Small talk Greetings, thanks, follow-up questions 4
+ more optimization, security, disk, network, fun 33

Proactive monitor

Background daemon that watches your system and pushes alerts without being asked:

  • CPU sustained >88% for 30s
  • RAM critical >93% → HOT strip (red, not chat spam)
  • CPU/GPU temperature spikes
  • Disk <8 GB free on any drive
  • Session uptime >12h reminder
  • New heavy process detected

Alerts go to the HOT strip (red) or TIP strip (yellow) depending on severity. RAM critical never appears as a chat message — only in the dedicated HOT indicator.


Bilingual design

Language detected per message, not per session. Mix Polish and English freely.

Polish diacritic normalization via ASCII-fold dual scoring: "dzieki" matches "dzięki", "wydajnosc" matches "wydajność".


Dependencies

hck_GPT is designed as part of PC Workman HCK and uses its data pipeline:

Dependency Used for
psutil Live process list, CPU/RAM/disk
pywin32 (WMI) CPU model, GPU name, VRAM, mobo, RAM speed
sqlite3 stdlib User knowledge base, historical stats
tkinter stdlib Chat panel UI
requests optional Ollama LLM API (local)

Standalone extraction as a pip-installable library is planned for a future milestone.


Version history

Version What changed
2.1.0 HOT strip for RAM alerts (no chat spam), tip_green advisory background, welcome_bg table styling, register_hot/clear callbacks, UZYTKOWNIK action tracking
2.0.4 Wave 2: 6 new intents (game_can_run, upgrade_feasibility, top_resource_hog, daily_ram_usage, battery_estimate, gaming_ram_usage). 82 intents total
2.0.0 Wave 1: 13 community-requested intents, Context Time-Windowing, No-AI-Slop fallback, Time-Travel Debugging, Micro-Benchmarking
1.7.x DeepMonitor integration, language sync, conversation flow, process library 373 entries
1.0.0 Initial: Hybrid Engine, 63 intents, Bordeaux Noir panel, proactive monitor, session memory

Part of PC Workman HCK

hck_GPT is the AI brain inside PC Workman HCK — a real-time Windows system monitor with 2.5D hardware map, DeepMonitor sensor table, startup/services manager, and time-travel diagnostics.

Marcin "HCK" Firmuga · GitHub · LinkedIn · MIT License

About

Local AI diagnostic assistant for Windows PC monitoring. 76 intents, hybrid rule and LLM engine, bilingual PL/EN, context time-windowing, anti-hallucination fallback, and time-travel debugging. Built from real community requests. Part of PC_Workman HCK.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages