Explore river records and run reproducible water analyses in your browser or in Python, with an agent on top.

• 37 global water-data sources behind one Pydantic schema: USGS, UK EA, Hub'Eau, GRDC, FAO, GEMStat, EU WFD
• Every river on a globe: ~6.8M GEOGLOWS reaches, 15-day forecasts, an A to D evidence ladder grading global models, flood layers back to 1990
• AI on the map: ask the map, explain this view, review your model results, run what-if scenarios
• Studio: a seven-role agent crew plans, runs, and signs off a study, returning Word, Excel, notebook, and engineering exports (HEC-RAS/HMS, SWMM, MODFLOW 6)
• MCP server for Claude or Cursor · HydroGym RL calibration environment
• 3,200+ tests · CAMELS-validated · 30+ outside contributors · Zenodo-archived

Open-source agriculture assistant you run on your own machine: a multimodal tool-calling agent over a deterministic, cited engineering core.

• LLM collects and routes; a validation gate guards a pure, tested trust zone that does the math, and every number in a reply is checked against what the engine computed
• Cited aquaponics and hydroponics design (10 fish, 34 crops), farm and crop records, and experimental field-irrigation checks (drip, sprinkler, soil water)
• Photos and voice notes on Telegram, WhatsApp, and the web; a cited table returns a ranked differential, never a verdict
• RAG measured, not asserted: hit 0.879 · recall 0.833 · MAP 0.604 · 8/10 off-topic refused; nine techniques benchmarked, four ship. Faithfulness ~0.8, citation accuracy 1.00, judge agreement reported too
• 1,770+ tests · advice-safety golden set in CI · agronaut eval · DPG compliance pack

A differentiable research twin of Taiwan's Choushui alluvial fan: pumping policy in, groundwater heads and land subsidence out, gated against held-out data at every stage.

• 174 monitoring wells, 116,769 metered pumps, 798 leveling benchmarks, 2012–2022, on a 1 km four-aquifer grid
• Full chain hindcast of subsidence: R² 0.639, RMSE 5.7 cm across 36 ensemble members; compaction column R² 0.652 out of fold
• Policy projections to 2032 (36 members × 3 policies × 21 years in ~50 min on one GPU) and a decision app with orthophoto and terrain
• Failures published, not hidden: on new 2023–2026 data the frozen forecast loses to seasonal climatology (RMSE 2.07 m vs 1.67 m), and every gate verdict is in the README

Speech recognition and translation for Mooré, a language of ~8M people and almost no training data.

• ASR corpus of 37,654 utterances / 85 h; fine-tuned MMS-1b reaches 16.8% WER / 4.3% CER on 564 held-out utterances, vs 31.1% zero-shot
• Curated 205,271 parallel pairs across 4 directions: LID-gated, fragment-filtered, decontaminated against FLORES-200
• LoRA fine-tunes of NLLB-200 600M and 3.3B on a single consumer RTX 4070
• Honest about what isn't solved: a native-speaker blind A/B puts translation at a coin flip, so it ships as unproven
• Three adapters on the Hub · Colab demo

A Claude Code plugin that turns the agent into a disciplined manuscript collaborator. Six strict modes (draft, review, revise, proofread, audit, format), Semantic Scholar MCP citation resolution with no API key, anti-fabrication guardrails, a number checker that traces every figure to its source, and a submission-ready .docx with a compliance report. Ships profiles for 10 journals (Hydrogeology Journal, J. Hydrology, WRR, HESS, Groundwater, IEEE TIM, and more) plus a generic quantitative-science profile. Install with one command: /plugin marketplace add Rekin226/paper-agent.