Hallucination aware framework for ranking LLM generated sports insights. 6 feature ScoreNet with PPO trained LLaMA reward models. Accepted at ECIR 2026 main track. Collaboration with Microsoft.
pytest for GPU kernels. Dtype-aware assertions, shape fuzzing, and CUDA benchmarking. Found 8 real bugs in Triton including an 83% error in layer norm.
Sub-15ms adversarial image sanitization for multimodal LLMs. C++17 SIMD core with AVX2/NEON dispatch, zero heap allocations in the hot path. pip install ready.
Benchmarking suite comparing three sparse attention methods as drop-in Llama replacements. Evaluated on MATH500, AIME, GPQA accuracy and latency from 1K to 128K tokens.
Explainable AI for exoplanet detection from Kepler/TESS light curves. CNN achieves 99.8% accuracy with per-prediction feature attribution. NASA Space Apps 2025.
Autonomous OSINT agent that fact-checks the internet in real time. NLP credibility scoring, multi-source synthesis, full Streamlit dashboard. Zero API keys.
Turn any scientific paper into structured, searchable experiment metadata in under a second. NER + 300 domain-specific regex patterns across 8 entity categories.
92.4% accuracy terrain classifier compressed from 150MB to 15.6MB. Compared 7 CNN architectures and 4 transfer learning approaches across 5 terrain classes. 21 stars.
BERT + MLP dual-model phishing detection with a Chrome extension. Real-time URL scanning via Flask API. Three training iterations with progressive feature engineering.
100% offline health risk intelligence. Ensemble ML (RF+LR) over 11 vital metrics with 7-day rolling analysis, PDF reports, and 6 Plotly visualization tabs. Zero cloud.