Trace
Track your job applications, see your priorities, find the skills you are missing.
A web demo that needs no installation · Explore it with sample data
Explore the screens · Built with AI · Technical details
The interface speaks Turkish. An English option is planned and will work
like the light/dark theme switch — see the roadmap.
The screenshots below are of the Turkish interface.
Trace at a glance
See scattered applications in one place. Spot what is coming up, compare stored
posting and CV assessments, and look at which skills to focus on. Urgency and
match are shown separately: an urgent job may not be your best match.
This is a demo. It uses a sample profile and anonymized records, and
connects to no personal account. Matches are computed from stored
assessments, and the courses are simulated. The images were captured from the
published demo on 20 September 2026.
Explore the screens
Click an image to open that screen of the app.
Try the app → · Read the product story →
Built with AI
Trace is an AI-assisted development portfolio that grew out of a real job
search. Atalay Denizer sets the product direction and the priorities. The
first version was built with Claude Code; Claude takes technical lead,
implementation and integration, while Codex contributes independent review,
testing and complementary development. Every change is reviewable through the
GitHub history and the pull requests.
The repository holds 10 agent roles and 7 skills. The posting analyzer and
the matcher have separate jobs; the two role advisors, one looking at the
profile and one at past outcomes, also draw on separate evidence. Agents
produce suggestions, a single designated writer stores the record, and the
schema and the arithmetic are validated by code.
These role definitions do not mean an AI service is wired into the web demo,
nor that model success has been measured. How the work is done, and the
contributions:
Claude instructions · Codex instructions · Collaboration · Contributing · Codex work hub
Technical details
The working demo, the architecture and the development notes are below. Open
whichever section you need.
| Capability | State today |
|---|
| Application list, filters, sorting, detail panel, two themes | A web interface running on demo data |
| CSV download of the visible applications | The active filter and on-screen order are preserved; the file is built in the browser |
| Urgency, match, reminders and eight analysis reports | Computed with the Python standard library; makes no model call |
| Interpreting a posting and deriving match dimensions | A Claude Code agent workflow; does not run automatically in the web demo |
| Gmail scanning | Requires a separately authorized session or automation |
| CV analysis | In a development environment with the right tool support; disabled in the public web demo |
| User accounts, saved data, connecting Gmail from the web | Not implemented |
| Training resources | Explicitly labelled simulation |
The account-based web application is separate work: web-live/
holds the Phase 1 scope and its development rules. The planned first release
covers login, CV upload, the onboarding flow and the application board. The
simplified application pipeline is defined as
saved → applied → interview → offer / rejected. Do not confuse that plan with
the current features of the static demo above.
Phase 1 scope · Web application rules
Authorized data collection and agent assessments feed the JSON records. The
Python core validates those records and computes urgency, match and the
reports. The builder embeds the demo data into the HTML template to produce the
web interface.
| Path | Responsibility |
|---|
data/ | Demo applications, profile and supporting catalogs across seven JSON files |
src/veri.py | Data source selection, schema validation, shared date and status definitions |
src/pipeline.py | Urgency, reminders, Markdown/text/CSV/JSON reports |
src/match.py | Summing match dimensions and assigning segments |
src/eslesme_kontrol.py | Schema and arithmetic check on an agent suggestion before it is stored |
src/insights.py | Eight analyses and the training plan |
src/anonimlestir.py | Deriving the demo data from real data |
src/dashboard.template.html | Source of the six-page interface |
src/build_dashboard.py | Data embedding and demo build |
tests/ | Python regressions and DOM behaviour tests |
.claude/ | Agent role definitions and task rules |
site/ | Overview page and the generated demo |
web-live/ | Phase 1 plan of the separate web application |
site/app.html, site/_artifact.html and reports/pano.html are generated
outputs. An interface change is made in the template and rebuilt.
In the design, indigo carries the brand, the purple tones carry match
quality, and red / amber / green carry process status. Theme and contrast
details are in the technical contract.
Trying the web demo needs no installation. To work on the code, Python 3.11+ is
enough; the computation core has no external Python dependency.
git clone https://github.com/atalay9807/trace-job-tracker.git
cd trace-job-tracker
python3 src/veri.py
python3 src/pipeline.py --format markdown
python3 src/match.py
python3 src/insights.py
python3 src/build_dashboard.py --site
python3 -m unittest discover -s tests -v
The interface development tests use Node 20+ and jsdom:
npm ci --ignore-scripts
npm test
git diff --check
These checks can run in the cloud or on
GitHub Actions. The
DOM tests do not verify real Gmail or model access, nor the visual layout in a
browser.
- Sample identity: Kerem Aydın is a demo profile; the project owner is
Atalay Denizer.
- Private data: real records are kept outside the public repository and
selected with
TRACE_DATA. If the selected external data is missing, the
code does not silently fall back to demo data. The data directory must be the
same for every script and for the relevant agents.
- Publishing limit: static HTML carries the data inside it and
authenticates nobody. A dashboard containing private data is never uploaded
to a public server;
--site accepts only the repository's demo data.
Personal application and email actions are not added to the public HTML.
- CSV: an export of the visible results. It is not a backup or import
format and does not modify the source JSON.
- Match:
match.py does not analyse the CV text; it sums the stored
dimensions. An unknown score stays null. Historical scores with no source
posting text do not count as verified model success.
- Rejection reasons: a gap inferred from the difference between a posting
and the profile does not prove the company rejected the person for that
reason.
- Replies and stages: the first reply is measured with
first_response;
last_contact cannot stand in for it. Without event history, current stages
are never presented as a sequential conversion rate. The total number of
postings ever shown to the user is unknown.
- Training: course titles, prices and the enrolment flow are examples. They
do not mean a real course service is connected.
The data flow and validation rules are in the
collaboration contract.
Moving to a multi-user product requires identity and data isolation, safe
record updates, server-side integrations and data management, all planned
separately. The acceptance conditions are in the
production readiness plan and in
ROADMAP.md; the limits of the first web release are in the
Phase 1 scope.
Atalay Denizer · AI-assisted development portfolio · GitHub profile
Open the app · Explore the site
Licensed under the MIT License · Code of Conduct
