Repository avatar
Other Tools
v1.0.0
active

tally

io.github.verkyyi/tally

Company and open-source numbers (valuation, revenue, prices, stars) with history and sources.

Documentation

Tally

Any number, with its history and its source. Tally keeps a public record of numbers about companies and open-source projects: funding, hiring, GitHub stars, downloads, batch sizes. Each number has its own page with a chart you can link to or embed, and every point links to where it came from.

What it does

  1. Pick who you watch. Companies, accelerators, sectors and investors. Nothing to type.
  2. Get a board. Funding, latest round, valuation, headcount, open roles and launches, from curated recipes.
  3. See what changed. Numbers are re-checked in the background; the Activity tab lists what moved.

Numbers live in one shared catalog, so a number one person tracks is there for everyone, with its history.

How a number is made

Every number is a small, open recipe:

  • Gather evidence: search, then read the pages.
  • Read the fields, guided by plain-language reading instructions in skills/.
  • Reduce to one number: read one value, count a list of entries, or add up, average or take the largest across them.

The recipes are in this repo: templates.json (19 curated recipes in company, accelerator, sector and investor packs) and skills/.

Three rules keep the numbers honest:

  • A value is accepted only if it can be read in the text it came from. Otherwise the card says "no value yet".
  • A count is a list you can open, one entry per link. It is never the size of a search result.
  • A funding round reported by two sites is one entry.

Search and page reading run on Telem, with Tavily pinned for finding entries. Extraction uses Claude.

Run

uv venv --python 3.11 .venv && uv pip install -p .venv/bin/python -r requirements.txt
cp .env.example .env            # TELEM_API_KEY, ANTHROPIC_API_KEY, LLM_MODEL
./run.sh                                  # http://127.0.0.1:8000

Stack: Python 3.11, FastAPI, SQLite, httpx, one static index.html. LLM: Anthropic API (LLM_MODEL, default claude-sonnet-5-5).

License

MIT. See LICENSE.

Local regression tests

.venv/bin/python -m unittest discover -s tests -v

Tests create temporary SQLite databases and mock network access, LLM responses and notifications. They do not need credentials or a running service.

Text-derived direct values pass a shared evidence review (evidence.py) using the existing LLM client. The runtime requires a quote on the cited fetched page, a matching entity and metric scope, an exact value, and an evidenced effective date (or an explicitly current undated source). Failed, ambiguous, bounded, approximate, or range values are saved in evidence_reviews, not published as exact numbers. The original quote, qualifier and bounds remain available there for later review; there is no review UI yet. LLM failure holds the candidate without replacing the last observation. Legacy headcount points without scope metadata require review before a changed value can produce a comparison.

Observations distinguish as_of (fact date) from collected_at (read time). Recipe repair, replacement and restoration must pass every required check. Monitor notifications trigger verified reads, rather than directly supplying direct values. Structured public API adapters keep their existing deterministic extraction path. No migration rewrites historical observations or sends corrections/alerts.

No installation packages available.

Remote