Sentinel

Build Windows apps in 13 seconds. All from your phone, through Telegram.

No expensive setup. Self-learning. Runs entirely on your machine.

You give it a one-line description in Telegram. Local Qwen writes the Python. PyInstaller bundles it. A double-clickable .exe lands on your Desktop. Thirteen seconds, end to end. Nothing in that loop talks to the cloud.

Most agent frameworks assume you have GPT-4 on tap or a 24 GB workstation. Sentinel was built on an RTX 3050. The whole design is shaped around one constraint: only one local model fits in VRAM at a time, so everything that doesn't have to be an LLM call gets pushed off the GPU.

Quick Start

git clone https://github.com/malqouqa92/Sentinel.git
cd Sentinel
.\setup.ps1          # 5-15 min: Python, Ollama, models, Claude CLI, creds
.\start_bot.ps1      # bot lives in Telegram from here

Then in Telegram:

/gwenask build me a stopwatch with start, stop, reset

You get back a recipe. Paste it as a new message:

/gwen STEP 1: run_bash command="..."
STEP 2: write_file path="..." content_b64gz="..."
...

Thirteen seconds later there's a Stopwatch_Setup.exe on your Desktop. Double-click. Done.

Why this is different

	Cloud chatbots	Server-side agents	Sentinel
Internet required for inference	Yes	Yes	No
API keys	Yes	Yes	None
Cost per use	Subscription	Per-token	$0
Hardware floor	None (cloud)	8+ GB VRAM	4 GB GPU, or just CPU + RAM
Builds a Windows .exe	No	Rarely	In 13 seconds
Improves over time	No	No	Self-learning
Privacy	Goes to provider	Goes to provider	Stays on your machine

What it can do

Build apps from a sentence /gwenask build me a stopwatch produces a STEP-numbered recipe. Paste it back. 13 seconds later there's a working .exe on your Desktop. No Visual Studio, no PyCharm, no electron-builder, no config files. Local Qwen 2.5 Coder 3B writes it; PyInstaller bundles it.	Hybrid by default. 80/20 split. Qwen 2.5 Coder 3B does roughly 80% of the work locally on a 4 GB GPU. The optional Claude CLI handles the remaining 20% as a ceiling for hard problems. It runs as a local subprocess against your existing Claude Code login, not an outbound API call. Sentinel itself never reaches the internet for inference.
Self-learning When /code succeeds, the recipe is stored as a pattern, replayed against a fresh tree, and graduated. Patterns Qwen can solve on its own get auto-pinned. After cold start, the same prompts skip Claude entirely and return in seconds.	Runs on a 4 GB GPU. Or any PC at all. Ollama supports CPU inference. The bot still works on a laptop with no discrete GPU, just slower. The hardware floor is whatever can hold ~3 GB of weights.

How it works

┌──────────┐     ┌────────────┐     ┌────────────┐     ┌──────────┐
│ Telegram │ ──► │   Router   │ ──► │ Job Queue  │ ──► │  Worker  │
└──────────┘     │  (no LLM)  │     │  (SQLite)  │     │ (async)  │
                 └────────────┘     └────────────┘     └────┬─────┘
                                                            │
                       ┌────────────┐     ┌─────────────┐   │
                       │  Skills    │ ◄── │ Orchestrator│ ◄─┘
                       │ (Python)   │     │ (dict map)  │
                       └────────────┘     └─────────────┘

The router parses /commands deterministically (no LLM call). The job queue is SQLite, which means it survives crashes for free. The worker holds an SQLite-backed GPU lock so only one inference happens at a time. When a skill needs an LLM, it hits Ollama first; if the problem is too hard for Qwen, the skill escalates to the local Claude CLI as a subprocess. Everything is logged with a per-request trace ID into logs/sentinel.jsonl.

Self-learning, in one screenshot

/code starts with Qwen. If Qwen can't solve it, the local Claude CLI fixes it, stores the recipe as a pattern, and Qwen gets graduated against the same problem on the next clean tree. After five graduations at 100% pass rate, the pattern auto-pins and short-circuits future requests.

Job-search pipeline

The first real test domain. Scrapes Indeed, LinkedIn, Glassdoor, and ZipRecruiter through python-jobspy. A pre-LLM title filter drops guaranteed-no postings before they cost you a Qwen call. A 5-dimension rubric (CV match, north-star, comp, cultural signals, red flags) scores survivors. A hard commute gate runs offline against a US zip dataset. Drops a CSV plus the top picks straight to your chat.

Hardware

GPU	Nvidia 4 GB VRAM minimum (tuned for RTX 3050). Or skip GPU entirely and run on CPU.
RAM	16 GB recommended. CPU-only mode wants more.
OS	Windows 11
Python	3.12
Disk	~5 GB after Qwen pulls

Model	Backend	Job
qwen3:1.7b (custom sentinel-brain Modelfile)	Ollama	Chat, classification, summaries
qwen2.5-coder:3b	Ollama	Code, JSON extraction
claude-cli (optional)	Local subprocess	Hard problems, teaching

Built on

• Ollama — local model runtime
• Qwen 2.5 / Qwen 3 — Alibaba's open-weight models
• python-telegram-bot — Telegram client
• PyInstaller — .py → .exe bundling
• python-jobspy — multi-board job scraper
• pgeocode — offline US zip → lat/long
• Claude Code — optional ceiling

Repo layout

core/                Config, telemetry, router, DB, worker, LLM, skills, agents
skills/              The actual capabilities (job_*, web_*, code_assist, file_io)
agents/              YAML files that wire skills into pipelines
interfaces/          The Telegram bot
workspace/persona/   IDENTITY, USER, MEMORY, SOUL, PROMPT_BRIEF, QWENCODER
tests/               One pytest file per phase, ECC-style
tools/               preload_kb.py and friends
scripts/             Install supervisor, emergency cleanup
main.py              The async event loop
setup.ps1            First-run installer
bootstrap.ps1        Single-file fetch-then-setup wrapper

Documentation

SETUP.md	Install paths, troubleshooting, uninstall
CLAUDE.md	Architecture, module map, design decisions
PHASES.md	Full change log, every phase, every footgun
PROMPT_BRIEF.md	Brief you upload to ChatGPT/Claude/Gemini to author /gwen recipes

Honest caveats

• The fast /gwenask path verifies cleanly on simple tkinter apps (stopwatch, counter, tic-tac-toe). Bigger asks like calculators and CSV viewers are designed and queued but not battle-tested in the automated harness.
• Job scrapers occasionally rate-limit. Indeed and LinkedIn are worst.
• Graduation tests inside /code add 30–60 seconds per successful pattern. Worth it (they catch silent failures), but it makes /code feel slow on the cold path.
• Windows-only. setup.ps1 is PowerShell. Linux/macOS support would mostly be a setup.sh plus path normalizations.

License

MIT. Do whatever you want with it. See LICENSE.