Open-source Python toolkit that unifies global water data under one API, then puts an agent on top of it.

• 29 collectors behind one Pydantic schema: USGS, UK EA, Hub'Eau, GRDC, FAO, GEMStat, EU WFD
• MCP server — Claude or Cursor gets stations, series, and cited flood frequency as real tools
• Studio: a six-role agent crew (scout, methodologist, analysts, critic, author) plans, runs, and returns a Word + Excel + notebook bundle
• HydroGym: gym-style RL environment for model calibration on real basins
• 1,000+ tests · CAMELS-validated · 24 outside contributors · Zenodo-archived

Multimodal tool-calling agent over a deterministic, cited engineering core. Zero proprietary APIs.

• LLM collects and routes; a validation gate guards a pure, tested trust zone that does the math
• Photos and voice notes on Telegram, WhatsApp, and the web — a guard strips any fabricated reading, and 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 queries refused. Nine techniques benchmarked, four ship, every verdict recorded
• Faithfulness + citation-accuracy eval; per-turn tracing that records shape, never content
• 1,235 tests · advice-safety golden set enforced in CI · DPG compliance pack

GPU physics-informed neural operators. One model replaces 61 hand-calibrated ones.

• Simulation KGE 0.754 vs 0.736 for 61 per-well calibrated ODEs (climatology 0.446)
• Generalizes to wells it never saw: leave-one-well-out 0.565, near the in-sample 0.591
• Forecast LSTM: 0.899 at 30 days vs 0.703 persistence, with ~90% calibrated intervals
• 14× bf16 GPU speedup · 6.7× parallel rollout · PhysicsNeMo port, bit-identical
• Negative results published, not buried: the PINN field, the adjoint solver, the subsidence coupling

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

• 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
• ASR corpus of 37,654 utterances / 85 h; Whisper-small at 34.1% WER, benchmarked against an MMS-1b zero-shot baseline
• Three adapters published on the Hub · BLEU / chrF++ on FLORES-200 devtest · Colab demo

A Claude Code plugin that turns the agent into a disciplined manuscript collaborator — five strict modes (draft, review, revise, proofread, audit), Semantic Scholar MCP citation resolution with no API key, anti-fabrication guardrails, and a clean .docx round-trip. Ships journal profiles for Hydrogeology Journal, JHRS, and IEEE TIM, plus a generic quantitative-science profile. Install with one command: /plugin marketplace add Rekin226/paper-agent.