PCSense - an AI technician that proves it fixed your PC before it tells you so

#ai #hacktoberfest #llm #tools

Hacktoberfest: Maintainer Spotlight

What we built

PCSense is a local-first AI technician for Windows 11. Instead of just showing raw
metrics (Task Manager) or blindly deleting files (traditional cleaners), PCSense
diagnoses why your PC is struggling, explains its reasoning in plain language,
proposes a fix, simulates the effect before anything changes, asks for
approval, and then verifies with a real before/after measurement whether the fix
actually worked.

"PCSense never says it fixed something until it has measured it."

Built for Hacktoberfest Hack Day 2026 (INIT Club x iDEA Club x MLH) by team YVL.

Key features

Natural-language request routing via Gemma 4, served locally through Ollama — with a deterministic keyword-baseline fallback if the model is unavailable
Multi-factor diagnosis: a primary cause, secondary contributors, and plain-language notes about metrics that are elevated but not actually a problem
A what-if simulator that predicts the effect of a fix (e.g. "RAM: 62% → ~41%") before you approve anything
A policy-gated safety core with adversarial tests: sandbox confinement, path-traversal and junction/reparse-point rejection, protected-process checks, and plan-hash binding so an approved plan can't be swapped out from under you
Reversible cleanup via quarantine (never a direct delete) and a full SQLite audit log
A faithfulness checker that verifies every number the AI states against real evidence
Battery, GPU (NVIDIA), network, driver inventory, and Windows Event Log ("PC Black Box") telemetry, a hardware upgrade advisor, a simple game-compatibility checker, and workload-mode-aware health scoring

How we built it

A shared, frozen pydantic contracts module so four people could build in parallel against one typed interface from minute one
A safety core written test-first and adversarially (junction traps, path traversal, protected paths/processes, plan-hash mismatches) before the real implementation
An orchestrator tying router → diagnosis → plan → approval → policy → executor → verify into one event stream a Streamlit UI consumes
Real integration work reconciling two tracks' independently-built (and differently shaped) agent-core implementations against the frozen contracts, without either side rewriting their own design
Honest degradation everywhere hardware telemetry isn't available (no NVIDIA GPU, no exposed thermal zone, etc.) instead of fabricating numbers

Tech stack

Python 3.11, Streamlit, Pydantic, psutil, pywin32/WMI, Ollama + Gemma 4 (schema-constrained
JSON output), SQLite, pytest.

Team

Shashank Sivakumar, Nikhil Sivakumar, Yashas Senthil Kumar, Hanshith Dhullipalla - Team YVL.

DEV Community