Work / No. 03

SafeTrace — local evidence review

A local-first prototype that reviews visual safety evidence on the user's own computer and links every finding back to its source frames. I contributed the integration that turns the team's analysis pipeline into a reviewable workflow.

My role: AI Solutions Intern · integration and connector work
Context: Internship prototype in a team repository
Status: Local-first prototype · not production
Period: 2026
Stack: Python · FastAPI · React · local vision models

What SafeTrace is

SafeTrace is a team prototype for reviewing visual safety evidence. It retrieves the frames likely to matter, runs vision models and deterministic checks on them, and organises the results as evidence a person can verify. The project's rules cover examples such as helmet, steering-wheel, phone and seatbelt conditions. The aim is review support, not automated judgement.

What I contributed

React/FastAPI integration

A React review interface connected to a FastAPI backend.

See Fig. 2

Local workflow hardening

Local job and batch processing that keeps one job's results separate from another's and can recover.

See Fig. 2

Evidence-review experience

Grouped findings, per-frame evidence and report export, so each finding ties back to its source frames.

See Fig. 2

Hosted-to-local connector

Lets the hosted interface talk to a SafeTrace backend running on the user's own computer.

See Fig. 1

I contributed these within a team repository. I didn't build the analysis pipeline: frame retrieval, the vision models and the rule layer are the team's work.

Where processing happensA hosted interface, with the analysis kept on the user's computer

The review interface can be served as a static page. The connector sends analysis requests from the browser over a loopback connection to a SafeTrace backend started on the same computer. The hosting provider serves page files only: footage and analysis results are not processed there.

My connector
Team-built and local components

Fig. 1 Where processing happens. Illustrative, simplified from the project's documented design.

WorkflowThe team's pipeline, made reviewable

The team's pipeline finds and checks frames. The integration work I contributed turns its output into grouped findings that a reviewer can verify against their source frames and export. It runs as local jobs behind the React interface and FastAPI backend.

Team-built
My integration contribution
A person decides

Fig. 2 The local review workflow. Illustrative; not a screenshot.

Technical challenges and decisionsIn the work I contributed

Decision	Why	Trade-off
01Connect a hosted interface to a local backend	The project is local-first. The connector lets people use a hosted interface while footage and results stay on their own computer.	The user has to start the backend locally, so there is no anonymous live demo.
02Treat the browser-to-local connection as a security boundary	A page loaded from the web is talking to software on the user's machine. The connector ships with automated tests for that boundary.	More setup and checks than a single-machine app.
03Tie every finding back to its source frames	A reviewer can verify a finding instead of trusting it.	Review takes a person's time; nothing is decided automatically.
04Harden local jobs and batches for isolation and recovery	Several files can be processed without one job's results mixing with another's, and work can recover after a failure.	More moving parts to run on one machine.

LimitationsThree boundaries

Technical

No approved public accuracy benchmark; the project includes an evaluation framework for human-labelled review
A prototype on an integration branch, not a production system
Analysis needs a SafeTrace backend running locally; the hosted page alone doesn't analyse anything

Ownership

A team repository; my part is the four contributions above
Frame retrieval, the vision models and the rule layer are the team's work

Data handling

Footage and analysis results stay on the user's computer
The hosted page serves the interface only
Rule checks are project rules, not certified safety policy

Stack · Python and FastAPI (local backend and jobs) · React with Vite and Tailwind (review interface) · local vision models (the team's pipeline)

Artwork

Artwork Abstract, not real footage or a system output.