yen-idhazh
A daily news digest that checks its own work.
Read it: https://miztiik.github.io/yen-idhazh/
idhazh is Tamil for a journal. yen-idhazh is "my journal".
Every morning it reads a curated set of public sources, writes a short summary of
what matters, and - the part that makes it different - scores each summary
against the article it came from and shows you the score. Where the check went
badly, the item says so.
Nothing runs when you read it. The page is plain files.
What you get
| |
|---|
| Five subjects, one page | AI, energy, business and economy, world, India. Ten to fifteen stories, not a firehose. |
| A confidence mark on every story | Each summary is machine-checked against its source article. A story that scored badly is labelled, not hidden. |
| The link, always | You get our summary and the address it came from. The article itself is never republished. |
| Two minutes, then out | A digest, not a feed. No infinite scroll, nothing to come back for. |
| Nothing follows you | No accounts, no analytics, no cookies, no third-party scripts. |
| A picture only when it earns one | Numbers become a chart. A process becomes a diagram. Everything else gets nothing. |
| Search on your device | Optional. One download, then the archive is searchable without anything you type leaving your browser. |
It is for someone who wants to know what happened in a few subjects, in two
minutes, from a page that says how much to trust it.
How it works
flowchart TD
subgraph build["Build time - Content refresh, four UTC runs daily"]
direction TB
A["<b>Collect</b><br/>read 138 live feeds<br/>rank by tier, repetition and age"]
B["<b>Extract</b><br/>pull readable text<br/>sanitise at the trust boundary"]
C["<b>Summarize</b><br/>Qwen3-8B on CPU<br/>output shape pinned by a schema"]
D["<b>Score</b><br/>faithfulness vs the source<br/>plus model-free counterweights"]
E["<b>Plan a visual</b><br/>a chart, or nothing<br/>the model picks numbers by index"]
F["<b>Assemble</b><br/>write the day's payload<br/>commit it to the repository"]
A --> B --> C --> D --> E --> F
end
F -->|"committed JSON + SVG"| G
subgraph read["Read time - a static page"]
direction TB
G["<b>Two kinds of page</b><br/>home, archive, scores prerendered<br/>a day page renders in the browser"]
H["<b>On-device search</b><br/>optional, reader-initiated"]
G -.->|"only if you click"| H
end
style build fill:#eef2ff,stroke:#4c6ef5
style read fill:#f0fdf4,stroke:#16a34a
The line between the two boxes is the whole design: everything expensive
happens before you arrive. The summaries are written and committed; a page
only reads them.
The one rule that shapes everything
The pipeline runs on a stock GitHub runner - 4 vCPU, no GPU, a 6-hour job cap.
That is not a budget to be raised; it is the platform. It decides which model can
be used, how work is sharded, and which features never ship.
When a measurement contradicts the design, the design changes. A laptop said the
8B model decoded 3.3x slower than the 4B, which nearly cost it the job. The
runner said 1.8x, so the 8B writes the summaries.
Where the stories come from
Sources are curated, not crawled. Every feed is listed by hand in
config/sources.json with a subject and a tier:
| Tier | What it is | Example |
|---|
| 1 | The institution that is the fact | a central bank, a statistical agency, a lab's own blog |
| 2 | Trade press that covers the beat daily | a specialist outlet, a wire's section feed |
| 3 | Community and aggregators | a forum, a link aggregator |
Ranking is arithmetic, not judgement: tier weight, times how many independent
sources carried the story today, plus a bonus for a watched entity. A story
three independent outlets carried is the day's story. No model is involved in
choosing.
Live feeds, counted 2026-08-24: ai 38, india 27, world 27, energy 24,
business-economy 22 - 138 in total.
A source that dies is retired in config, never deleted - deleting an id would
break every payload that referenced it. See
docs/architecture/sources/discovery.md.
Documentation
Start here, in this order:
Local development
python -m venv .venv
.venv/bin/pip install -e ".[dev]" # .venv/Scripts/pip on Windows
.venv/bin/python -m pytest # tests
.venv/bin/python -m ruff check . # lint
.venv/bin/python -m mypy # strict type check
Production runs in GitHub Actions; local model inference is not supported.
Use the Pipeline tests workflow
to exercise the production stages with configured article cases. Model weights
and llama.cpp binaries are downloaded in CI and never committed.
Repository layout
backend/ produces and frontend/ publishes. They meet only through committed
files and the schemas generated from backend/idhazh/contracts/ - never through
a running service. Every directory is explained in
docs/reference/repository-layout.md.
See also
