2026 · ACTIVE · BUILD

ContextGuard

A webcam that reasons about context instead of shouting at motion

Python
YOLO26n
ByteTrack
SQLite
FastAPI
Streamlit
SmolLM2
Docker

ContextGuard turns an ordinary laptop webcam into a context-aware physical-security sensor. It detects and tracks people, reasons about where they are and how long they linger, scores the situation with an explainable risk engine, records structured events to a local SQLite store, and writes a grounded natural-language narrative for each one.

The design is built on three commitments. Context over detection: an alert is a function of zone, time of day, dwell time, repeat visits and transition patterns, not just the existence of a bounding box. Explainable, evaluable risk: three risk approaches share one feature set and one output shape so they can be compared fairly against three naive baselines. Privacy by construction: anonymous tracking is the default, no cloud LLM sits in the default path, and retention is actually enforced.

It is scoped honestly — a dorm room, a server rack, a home office. Not a surveillance product.

Problem

Most "AI security camera" projects stop at *a person is on screen* and fire an alert. That is noisy, context-free, and — because it usually ships frames to a cloud model — a privacy problem.

Question

Can a single laptop webcam produce alerts that a human could independently verify with a clock and a stopwatch, without a GPU, without the cloud, and without retaining video?

Approach

Separate detection from judgement. Detection is a tracker. Judgement is a swappable risk engine over verifiable context signals — which zone, what time, how long, how often, which transition. The narrator then describes the event in plain English and a hallucination checker refuses any sentence the event record does not support.

System

Open any stage to see what it does.

FRAME — A single webcam frame. CPU only. Nothing leaves the machine on the default path.

DETECT + TRACK — YOLO26n finds people; ByteTrack assigns a temporary identity. Anonymous tracking is the default — the ID is not a person, it is a thread.

CONTEXT — User-drawn zones, dwell time, loitering, repeat visits, after-hours, abnormal zone transitions. Every signal is one a human could check with a clock.

RISK ENGINE — Three interchangeable approaches — hand-written rules, model-derived weights, logistic regression — sharing one feature set and one output shape, so they can be compared against naive baselines rather than trusted.

EVENT STORE — Structured events in local SQLite. Retention is enforced by a scheduled purge, not by a config value nobody reads.

GROUNDED NARRATION — A small language model writes the event up. A hallucination checker then rejects any claim the structured event does not support. The narrator reports observations, never verdicts.

SURFACES — Streamlit dashboard for watching live, FastAPI REST for integration, plain-English query over the event history, Prometheus metrics for operations.

Demo

Streamlit dashboard runs the live camera view, zone editor, and event feed. The FastAPI service exposes the same event stream for integration. Both run from a single machine with no GPU.

Result

Runs CPU-only. No GPU and no mandatory cloud dependency on the default path.
Anonymous tracking by default; no thumbnails stored unless explicitly opted into; retention enforced by a scheduled purge.
The repository reports 81 passing tests and CI running on every push.
Project status is declared alpha by its own README — a research and small-deployment tool, not a product.

81 tests passing — Reported by the repository CI badge and test suite.

CPU-only inference — Stated hardware constraint of the project; no CUDA path exists.

Lesson

The hallucination checker was the part I expected to be a formality and it was not. Writing a checker forces you to define what the system is actually entitled to claim — and it turned out the honest answer was much narrower than the first draft of the prose. The narrator got quieter and the project got more trustworthy.

Source

github.com/BakulBd/ContextGuard ↗

Connected

CausalGuard — Both are AI and SYSTEMS work, and both are built with Python, FastAPI, Docker.
CodeRelay — Connected through AI and SYSTEMS.
TeleFlow — Both are SECURITY and SYSTEMS work, and both are built with FastAPI, Docker.
Epistemic Guard — Connected through AI and SYSTEMS.