Search awesome repositories

Entry to the Predictive AI Evaluation Competition at NeurIPS 2026 (Stanford AIMS): Beta-Binomial shrinkage predictor, independent audit, reproducible protocol and public-data findings.

ARC-AGI-2 contest implementation: CHIMERA symbolic candidates, reproducible ablations and clearly separated public evaluation and official scores.

Offline ARC-AGI-3 graph exploration agent with public-game evidence and reproducible evaluation.

Use the cheapest NVIDIA Nemotron that actually delivers: bounded agent tasks on Nebius Token Factory, deterministic checks and an auto-tier ladder (Lightning → Super → Ultra).

Emulating the ED global land-ecosystem model for the IEEE BigData Cup 2026: the anchor that beat every trained model, and the measured rejection of H1.

Reliability by construction for long-horizon healthcare agents on χ-Bench (IEEE Big Data Cup 2026, Challenge 07): contract-checklist hooks, critic panel and closure gate on the official Claude Code harness.

ClimX distributional climate emulator adaptation with bounded-memory prediction, upstream attribution and exact numerical-equivalence checks.

Structure-to-network developmental neurotoxicity triage agent with calibrated uncertainty (SciencePal AI4Science Hackathon 2026)

NAVAJO privacy gateway for JupyterHealth: reproducible synthetic-data benchmarks, clinical-text masking and explicit residual-risk reporting.

Pseudonymise clinical notes on the clinician's computer before any AI assistant, restore the answer. Pre-registered evaluation on MEDDOCAN (prototype, MIT Solve 2027).

MetFrag-assisted, leader-safe ranking of natural-product structures from MS/MS spectra — Enveda CASMI 2026 (Kaggle) entry with pre-registered experiments and animated figures

Evidence-linked funding-call review with Meterless H-MEM. Submitted 23 Oct 2026; dated Zetaris/Nightshift live-run evidence, synthetic sample mode and explicit operational limitations.

Controlled Triton kernel optimizations for FlagOS Track 1: three implemented operators, measured tiling interventions, full raw timings and correctness gates.

MiniCPM5-2B competition benchmark harness, numerical gates and three fixed experiment directions; official hardware evaluation pending.

Track 03 (Sleep onset) of the EEG/EMG Foundation Challenge 2026: a control that beats the published baselines, and a measured negative result for the Holographic-Reservoir

Online holographic-reservoir forecasts and gated PI control for the NeurIPS Smart Buildings Challenge; official simulator measurements and frozen prospective validation.

Autonomous, evidence-first engineering system for the BuildArena S02 (Besiege) Human-AI Engineering Challenge: rules-validated scorer, orbit/landing gates, isolated Autopilot agent run, submission packager.

Closed-loop tracking of actionable screening results with NAVAJO privacy for AI drafting. Prototype for the Future Health Digital Innovation Challenge (Challenge 05).

Rule-compliant Overture Maps coverage-gap estimator for 9,379 US tracts + a provenance bias discovery (Zindi x Humane Intelligence Bias Bounty)

Agent-native research lab: AI agents research, review each other and publish (P2PCLAW), with a pre-registered ablation of whether the peer review helps. Agent-Native Hackathon 2026.

Let an AI agent run PayPal payments without ever holding the credential, choosing the payee, or raising the amount. 450-run bench, MIT.

Airline satisfaction: nested GBDT ensembles, Darwin-Cage residual diagnostics, complete experiment records and attributed public prediction sources.

Fold-safe ensembling to the honest ceiling of Kaggle Playground S6E9, and a Darwin-Cage residual search with a power analysis. Nested OOF 0.946420; 47,679 evolved programs, 0 discoveries.

A local Gemma 4 jury reads the first 10 steps of a SWE-agent run, predicts failure, and lets a fixed budget resolve 29% more issues. Pre-registered study, Gemma 4 Paper Track.

TrafficFlowBench research implementation: conservation features, queue forecasting, reservoir ablations and reproducible scientific reporting.

Per-pixel voting field for muscle architecture (PA, FL, MT) from B-mode ultrasound - UMUD Challenge

Explainable agent that re-checks a grant compliance register against a rewritten federal regulation, using one federated Zetaris view over PostgreSQL, files and the live eCFR API.

zsh plugin and Claude Code agent skill to open a new Warp tab with a pre-loaded Claude session

Grafana-native LLMOps dashboards for FinOps, agents, quality and AI governance, generated and visually verified.

Multi-cloud cost management MCP server (Azure + GCP). Read-only access to spending, anomalies, forecasts, budgets, and guided FinOps workflows. Ask your AI about your cloud bill.