Which power projects will actually get built, and why.
▶ Try it now — no setup needed: https://queuescore.tech/
Winner, Candid Intelligence Hackathon (Track 1, "Project Radar") · Aug 2026
Most projects that enter the interconnection queue never get built, and the same real-world project hides under different LLC names in different agency systems. QueueScore stitches Texas's two live public paper trails — the ERCOT interconnection queue and TCEQ air permits — into one map, scores every project's completion probability, and drafts the brief for the call you should make about it.
3,700+ live filings on one map, colored by status — with the stitched lens showing projects confirmed across both sources. Click a pin and every panel follows.
Every ERCOT project gets a completion probability — an XGBoost model trained on ~15 years of national queue outcomes (LBNL "Queued Up") — with a SHAP breakdown of what drove it, and a deterministic stage ladder with the evidence for each call. Nothing is a black box.
ERCOT filings carry no coordinates — TCEQ permits do. When a project is stitched, it inherits its permit's exact location: flip to Site view and inspect the actual site from orbit.
One click drafts a seven-part origination brief (verdict, why now, snapshot, who, angle, evidence, gaps). Free-form questions get answers grounded in the record's actual fields — including what the filing doesn't show.
- Cross-source stitching, precision-first. County gate → name similarity → Claude adjudication of ambiguous pairs → an Opus second-opinion pass that vetoes weak matches. Every link carries a written reason; ambiguous pairs stay unlinked.
- Rules where rules win, models where they don't. Deterministic gates and
stage ladders do the auditable work; the LLM only gets genuine judgment
calls. The model never sees outcome-encoding columns — leakage is banned in
features.pyand enforced by tests. - Graceful degradation. Snapshot cache when offline, full map/scores/links with no API key. The demo can't be killed by wifi.
Stack: Python · XGBoost · SHAP · Streamlit · Plotly · gridstatus · Anthropic API.
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
streamlit run src/app.pyThe first run pulls both sources live and caches snapshots; afterwards it
works fully offline. The trained model and match table are committed, so
scores and stitched links need no API key — add ANTHROPIC_API_KEY to .env
for briefs, Q&A, and match adjudication. Rebuild the match table with
python -m src.resolve (add --second-opinion for the Opus audit pass).
Tests: pytest (29, all offline).
More docs: SETUP.md · DATA.md · SOURCES.md (incl. the reverse-engineered TCEQ endpoint) · DEPLOY.md · CHARTER.md
Built in a day by three people:
Or find us via the Get in touch button in the app itself.
Training data: LBNL Queued Up (CC BY 4.0). Sources: ERCOT interconnection queue (via gridstatus) · TCEQ Permit Search.