Curious mind building production-grade AI/ML systems. Focused on novel model architectures, uncertainty quantification, and efficient training strategies.

• body-blind-action-interface - Body-blind action interface that deletes embodiment information from the action space for zero-shot transfer across robot arms
• cross-embodiment-transfer-benchmark - Benchmark and analysis showing morphological similarity and data diversity do not govern cross-embodiment transfer
• cross-embodiment-measurement-standard - Enforced measurement standard and toolkit for evaluating cross-embodiment transfer claims
• symbolic-interface-transfer - Morphology-invariant symbolic interface for multi-step cross-embodiment policy transfer
• sufficiency-invariance-bound - Theory and measurements for the sufficiency-invariance bound on shared robot control representations
• mjx-crossmorph-locomotion - GPU-batched MuJoCo MJX pipeline for training and transferring joint-space quadruped locomotion across morphologies, built from scratch
• dreamerv3-robotic-control - DreamerV3 (29.5M params) for robotic control with RSSM world model, categorical latents (32x32), straight-through actor gradients, trained entirely in imagination on DMControl walker-walk
• action-conditioned-video-worldmodel - CogVideoX-5B fine-tuned with LoRA (rank 16) and novel action cross-attention adapter (95.2M trainable) for action-conditioned video prediction
• llm-planning-with-world-models - LLM planning with learned world models: Claude proposes action plans, DreamerV3 RSSM simulates outcomes, value head scores trajectories
• offline-robotic-manipulation-curriculum - Offline RL with curriculum learning for robotic manipulation

• ConstructAI-oss - AI-powered construction management platform driven by 11 coordinated agents, a local-first LLM gateway (vLLM + Ollama), construction-specialized RAG, and YOLOv8 + ViT computer vision
• llm-planning-with-world-models - LLM planning with learned world models: Claude proposes action plans, DreamerV3 RSSM simulates outcomes, value head scores trajectories
• Throughline - AI construction scheduling and project-controls platform: a deterministic CPM/DCMA/EVM engine with an LLM copilot layered around it as decision-support
• chw-companion - Offline multimodal maternal-health triage with Gemma 4, emitting structured on-device triage decisions (record_vitals, flag_danger_sign, recommend_action, schedule_followup) in a single forward pass

• adaptive-contrastive-curriculum-for-multitask-knowledge-transfer - Multi-task learning with adaptive curriculum and contrastive loss on MMLU
• adaptive-curriculum-learning-for-domain-transfer-in-llm-evaluation - Curriculum learning for LLM domain transfer on MMLU
• adaptive-difficulty-curriculum-mmlu-with-uncertainty-pacing - Uncertainty-paced curriculum learning for MMLU evaluation
• adaptive-reasoning-chain-distillation - Chain-of-thought reasoning distillation with adaptive selection
• CascadeExit-Research - Adaptive early-exit speculative decoding for LLM inference. 1.76x speedup on Llama-3.2-3B with SwiGLU exit adapters trained on WikiText-103
• contrastive-curriculum-mmlu-with-adaptive-difficulty-sampling - Contrastive learning with adaptive difficulty sampling on MMLU
• emergent-reasoning-via-grpo - Emergent chain-of-thought reasoning via GRPO on Qwen2.5-3B
• instruction-complexity-aware-lora-routing - Mixture-of-LoRA-experts with dynamic instruction routing on Alpaca
• meta-adaptive-prompting-for-multitask-reasoning - Meta-adaptive prompting strategies for multi-task reasoning
• metacognitive-confidence-calibrated-mmlu-reasoning - Metacognitive confidence calibration for MMLU reasoning
• MoLE-LoRA - Mixture of LoRA Experts with BERT-tiny router on Llama-3.2-3B. 73% memory savings, evaluated on MMLU/GSM8K/HellaSwag/ARC
• wiki-longctx-ssm-vs-transformer - Comparative analysis of SSM (Mamba) vs Transformer architectures for long-context modeling

• adaptive-retrieval-qa-with-answerability-calibration - Retrieval-augmented QA with answerability calibration on SQuAD 2.0
• contrastive-hierarchical-retrieval-augmented-qa-with-difficulty-routing - Contrastive hierarchical retrieval-augmented QA with difficulty routing
• contrastive-qa-verifier-with-adversarial-unanswerable - Dual-encoder QA verification with adversarial unanswerable detection
• genre-adaptive-nli-summarization-validator - Cross-genre NLI-based summarization validation on CNN/DailyMail + MultiNLI
• hierarchical-contrastive-qa-with-adversarial-unanswerable-detection - Hierarchical span prediction with contrastive learning on SQuAD 2.0
• legal-clause-risk-scorer - DeBERTa-v3 multi-task contract clause risk assessment on CUAD + LEDGAR
• nq-structure-aware-qa - Structure-aware QA on Natural Questions
• wiki-evidence-hallucination-detector - Evidence-based hallucination detection using Wikipedia

• adaptive-dpr-nq-retriever - Adaptive DPR retriever on Natural Questions
• adaptive-hybrid-msmarco-retrieval - Adaptive hybrid sparse-dense retrieval on MS MARCO
• adaptive-late-interaction-retrieval - ColBERT-style late-interaction retrieval with hard-negative mining on MS MARCO
• adaptive-msmarco-dpr - Adaptive hard-negative DPR for MS MARCO passage retrieval
• aurora-hybrid-msmarco - Query-adaptive sparse-dense hybrid retrieval on MS MARCO
• contrastive-retrieval-reranking-with-adaptive-hard-negative-mining - Contrastive reranking with adaptive hard negative mining
• entropy-colbert-msmarco - Entropy-guided ColBERT late-interaction retrieval on MS MARCO
• marco-matryoshka-retrieval - Matryoshka dense retrieval on MS MARCO
• matryoshka-dense-retrieval-msmarco - Matryoshka representation learning for dense retrieval on MS MARCO
• msmarco-adaptive-hybrid-retrieval - Adaptive hybrid retrieval for MS MARCO passage ranking
• msmarco-budgeted-neural-retrieval - Budget-aware neural retrieval on MS MARCO
• msmarco-caldera-retrieval - CALDERA retrieval pipeline on MS MARCO
• msmarco-calibrated-reranker - Calibrated neural reranking on MS MARCO
• msmarco-colbert-curriculum - ColBERT with curriculum learning on MS MARCO passage retrieval
• msmarco-colbertv2-pruning - ColBERTv2 with token pruning for efficient MS MARCO retrieval
• msmarco-distilled-hybrid-retrieval - Distilled hybrid retrieval for MS MARCO passage ranking
• msmarco-hybrid-retrieval - Hybrid BM25 + dense retrieval on MS MARCO
• msmarco-hybrid-retrieval-lab - Experimental hybrid retrieval on MS MARCO
• msmarco-matryoshka-curriculum-retrieval - Matryoshka curriculum retrieval on MS MARCO
• msmarco-qatp-colbert - Query-adaptive token pruning for ColBERT on MS MARCO
• q-pact-msmarco-retrieval - Q-PACT query-adaptive token pruning for late-interaction retrieval
• so-dense-retrieval-hard-negatives - Dense retrieval with hard negative mining on Stack Overflow
• so-duplicate-retrieval-lab - Duplicate question retrieval on Stack Overflow
• stackdup-dual-encoder-retrieval - Dual-encoder duplicate detection on Stack Overflow
• wikilink-graph-retrieval - Graph-enhanced retrieval using Wikipedia link structure
• wikiqa-distilled-hybrid-retriever - Distilled hybrid retrieval for WikiQA answer selection

• 3dgs-depth-regularized - 3D Gaussian Splatting with depth regularization on Mip-NeRF 360
• adaptive-noise-schedule-diffusion-with-clip-guidance - Adaptive noise schedule diffusion with CLIP guidance
• coco-rag-captioning - Retrieval-augmented image captioning with BLIP and LoRA on COCO
• hierarchical-diffusion-refinement-with-aesthetic-consistency - Multi-stage diffusion with aesthetic consistency loss for coherent multi-resolution generation
• hierarchical-diffusion-rlhf-caption-refinement - Hierarchical diffusion with RLHF caption refinement
• hierarchical-diffusion-with-preference-guided-refinement - Hierarchical diffusion with preference-guided refinement
• iterative-refinement-text-to-image-with-preference-guided-diffusion - Iterative text-to-image generation with preference-guided diffusion
• multi-objective-preference-diffusion-for-controlled-text-to-image - Multi-objective preference diffusion for controlled text-to-image
• multimodal-contrastive-captioning-with-preference-aligned-generation - Multimodal contrastive captioning with preference alignment
• preference-guided-diffusion-steering - Preference-guided steering for diffusion models
• preference-guided-image-captioning-alignment - CLIP + GPT-2 with DPO preference alignment on COCO and UltraFeedback
• progressive-diffusion-distillation-with-adaptive-step-curriculum - Progressive diffusion distillation with adaptive step curriculum
• rectified-flow-generation - Rectified Flow (1-RF) for unconditional CIFAR-10 generation. FID-50K of 5.39 with 100 Euler steps, surpassing published 1-RF baseline of 6.18

• adaptive-panoptic-segmentation-with-cross-scale-attention-fusion - Panoptic segmentation with cross-scale attention fusion
• adaptive-quantization-aware-serving-with-dynamic-precision - Dynamic INT8/FP16/FP32 precision switching with quantization-aware training on CIFAR-10
• adaptive-scale-aware-panoptic-segmentation-with-boundary-refinement - Scale-aware panoptic segmentation with boundary refinement
• adaptive-scale-fusion-urban-segmentation - Adaptive scale fusion for urban scene segmentation
• adverse-weather-semantic-segmentation-robustness-benchmark - Semantic segmentation robustness benchmark under adverse weather
• aerial-building-footprint-segmentation-with-boundary-refinement - U-Net++ with EfficientNet-B4 for aerial building footprint segmentation
• adversarial-uncertainty-calibration-for-medical-diagnosis - Adversarial uncertainty calibration for medical image diagnosis
• coco-boundary-calibrated-instance-seg - Boundary-calibrated instance segmentation on COCO
• coco-detr-lora - LoRA-adapted DETR for object detection on COCO
• coco-tailcal-detr - Tail-class calibrated DETR on COCO
• coco-uncertainty-selftrain - Uncertainty-guided self-training on COCO
• dacl10k-bridge-defect - 3-model UPerNet ensemble (DINOv3-L + EVA-02-L + ConvNeXt-L) for multi-label bridge-defect segmentation on dacl10k. val mIoU 0.476, +0.062 over the baseline (winning 17/18 classes)
• densenet-multilabel-chest-xray-pathology-detection - DenseNet-121 multi-label chest X-ray pathology detection on NIH ChestX-ray14
• hierarchical-attention-fusion-urban-scene-parsing - Hierarchical attention fusion for urban scene parsing
• pet-breed-uncertainty-aware-classifier - EfficientNet-B0 with MC Dropout uncertainty on Oxford-IIIT Pet (7,393 images)
• progressive-uncertainty-aware-pretraining-medical-xray - Progressive uncertainty-aware pretraining for medical X-ray classification
• rf-detr-detection - RF-DETR real-time object detection: 56.4 AP on COCO val2017 at 163 FPS
• s2ds-defect-segmentation - DINOv3-L/16 fine-tuned for structural-defect segmentation on S2DS (7 classes). test mIoU 0.844, +18 mIoU over the dataset paper baseline and the best of a 16-method benchmark
• tailforge-lvis-ovseg - Frequency-aware open-vocabulary segmentation with logit adjustment on LVIS
• uncertainty-aware-chexpert-diagnosis - Uncertainty-aware diagnosis on CheXpert chest X-rays
• uncertainty-aware-hierarchical-medical-diagnosis-with-credal-sets - Hierarchical medical diagnosis with credal set uncertainty
• urban-occlusion-aware-depth-estimation - Occlusion-aware monocular depth estimation for urban scenes
• urban-occlusion-aware-segmentation - Occlusion-aware urban segmentation with SegFormer + DeepLabV3+ ensemble

• rohbau3d-benchmark - First published semantic-segmentation benchmark on Rohbau3D — 504 terrestrial-LiDAR shell-construction scans (18 classes); Concerto-large (PT-v3m2) anchor at 0.584 mIoU
• scannetpp-dinov3 - DINOv3-injected-into-3D semantic segmentation: frozen DINOv3 ViT-L features injected at every decoder block of a Point-Transformer-V3 backbone on ScanNet++ (0.514 mIoU, near-SOTA)
• slabim-scan-to-bim - Concerto-large (PT-v3m2) point-transformer fine-tuned for scan-to-BIM semantics (wall/floor/door/column) on SLABIM, beating FM-Fusion 82.5 vs 58.3 mIoU

• adaptive-spectral-attention-heterogeneous-molecular-property-prediction - Spectral attention for heterogeneous molecular property prediction
• adaptive-spectral-message-passing-for-molecular-scaffold-learning - Spectral message passing for scaffold-based molecular learning on BBBP
• adaptive-spectral-rewiring-for-heterogeneous-molecular-graphs - Spectral rewiring for heterogeneous molecular graphs
• adaptive-spectral-temporal-gnn-for-molecular-toxicity-prediction - Spectral-temporal GNN for toxicity prediction on Tox21 with curriculum learning
• hierarchical-attention-pooling-for-molecular-scaffold-transfer - Hierarchical attention for molecular scaffold transfer on MoleculeNet BBBP
• hierarchical-message-passing-with-curriculum-weighting-for-molecular-toxicity - Hierarchical message passing with curriculum weighting for toxicity
• molecular-scaffold-aware-multi-task-toxicity-prediction - Scaffold-aware GCN with attention pooling on Tox21 (7,823 molecules)
• molecular-scaffold-transfer-learning - Transfer learning across molecular scaffolds
• pcqm4mv2-3d-prior-distillation - Distilling 3D quantum priors into 2D graph transformers on PCQM4Mv2
• spectral-temporal-curriculum-molecular-gap-prediction - Spectral graph wavelets with curriculum learning on PCQM4Mv2 (3.7M molecules)
• spectral-temporal-curriculum-molecular-gaps - Spectral-temporal curriculum learning for molecular gap prediction
• temporal-drug-interaction-prediction-with-heterogeneous-molecular-graphs - Temporal drug interaction prediction with heterogeneous graphs

• mamba-genomic-classification - Mamba SSM for genomic sequence classification: 18-task benchmark vs CNN and Transformer

• adaptive-feature-importance-reweighting-credit-risk - Adaptive feature importance reweighting for credit risk
• adaptive-feature-interaction-learning-with-dynamic-gating - Hierarchical feature interaction learning with adaptive gating
• adaptive-feature-interaction-networks-with-dynamic-gating - Feature interaction networks with dynamic gating for tabular data
• adaptive-feature-selection-with-uncertainty-aware-ensemble-for-fraud-detection - Adaptive feature selection with uncertainty-aware ensemble for fraud detection
• adaptive-gating-network-for-heterogeneous-tabular-ensembles - Gating network for heterogeneous tabular ensembles
• adaptive-gradient-boosting-with-dynamic-feature-synthesis - Gradient boosting with dynamic feature synthesis and meta-learning
• adaptive-threshold-calibration-fraud-detection-with-cost-sensitive-curriculum - Cost-sensitive curriculum learning for fraud detection threshold calibration
• adaptive-threshold-ensemble-with-uncertainty-calibration-for-fraud-detection - Threshold ensemble with uncertainty calibration for fraud detection
• fairness-aware-income-prediction-with-constraint-optimization - Fairness-constrained LightGBM with Optuna on UCI Adult Census
• medicaid-spending-analysis - Statistical and ML analysis of HHS Medicaid provider spending data
• multi-horizon-credit-default-forecasting - Multi-horizon credit default forecasting
• neural-two-tower-product-recommendation-with-contrastive-learning - Neural two-tower product recommendation with contrastive learning
• temporal-credit-degradation-detector - Temporal credit degradation detection
• temporal-drift-aware-fraud-detection-with-adversarial-validation - Drift-aware fraud detection with adversarial validation

• adaptive-hierarchical-reconciliation-with-attention-pruning - Hierarchical reconciliation with attention-based pruning
• adaptive-horizon-meta-learning-hierarchical-forecasting - Horizon meta-learning for hierarchical forecasting
• adaptive-traffic-anomaly-forecasting-with-uncertainty-quantification - Traffic anomaly forecasting with uncertainty quantification
• hierarchical-attention-reconciliation-retail-forecasting - Attention-based reconciliation for retail forecasting
• hierarchical-demand-forecasting-with-reconciliation - Hierarchical demand forecasting with reconciliation
• hierarchical-demand-reconciliation-with-temporal-anomaly-feedback - Demand reconciliation with temporal anomaly feedback
• hierarchical-forecast-reconciliation-with-uncertainty-quantification - Forecast reconciliation with uncertainty quantification
• hierarchical-traffic-demand-forecasting-with-graph-informed-reconciliation - Graph-informed reconciliation for traffic demand forecasting
• multi-horizon-probabilistic-energy-forecasting - Probabilistic energy forecasting with PatchTST, TFT, and N-BEATS

• adaptive-chunk-conformer-rnnt - Adaptive-chunk Conformer RNN-T for streaming ASR on LibriSpeech
• adaptive-streaming-asr-librispeech-musan - Noise-robust streaming conformer ASR with MUSAN augmentation
• conformer-ctc-librispeech-rigorous - Rigorous Conformer CTC with SpecAugment curriculum on LibriSpeech 960h
• cross-corpus-speech-emotion-recognition-with-wav2vec2 - Wav2Vec2 speech emotion recognition on RAVDESS and CREMA-D
• cross-lingual-phoneme-aware-speech-enhancement-with-adaptive-masking - Cross-lingual phoneme-aware speech enhancement
• cross-lingual-phonetic-adapter-speech-recognition - Cross-lingual phonetic adapter for speech recognition
• ctc-selftrain-librispeech - CTC self-training on LibriSpeech
• librispeech-mrm-wav2vec2 - Masked representation modeling with wav2vec2 on LibriSpeech
• librispeech-semisup-conformer - Semi-supervised Conformer CTC with pseudo-labeling on LibriSpeech
• librispeech-uncertainty-conformer - Uncertainty-aware Conformer CTC ASR on LibriSpeech
• librispeech-uncertainty-consistency-asr - Consistency-regularized uncertainty-aware Conformer ASR on LibriSpeech
• mls-adapter-asr - Adapter-based multilingual ASR on MLS
• multi-scale-temporal-fusion-speaker-diarization-with-uncertainty - Multi-scale temporal fusion speaker diarization with uncertainty
• multi-speaker-diarization-with-adaptive-boundary-refinement - Multi-speaker diarization with adaptive boundary refinement
• noise-aware-conformer-asr - Noise-aware conformer for robust ASR
• phoneme-aware-curriculum-diarization-with-adaptive-merging - Phoneme-aware curriculum diarization with adaptive merging
• phoneme-aware-multilingual-asr-with-adaptive-layer-fusion - Phoneme-aware multilingual ASR with adaptive layer fusion
• progressive-multi-task-speaker-diarization-with-asr - Progressive multi-task speaker diarization with joint ASR
• robust-wav2vec2-consistency - Consistency-regularized wav2vec2 for robust ASR
• ssl-asr-robust-librispeech - Self-supervised learning for robust ASR on LibriSpeech
• streamconformer-distill-asr - StreamConformer distillation for efficient ASR
• streaming-asr-adapters-librispeech - Adapter-based streaming ASR on LibriSpeech
• streaming-asr-conformer-librispeech - Streaming conformer ASR on LibriSpeech
• streaming-conformer-asr - Streaming conformer architecture for ASR
• streaming-conformer-distill-librispeech - Knowledge distillation for streaming conformer ASR
• ucc-librispeech-longform-asr - Unified context carry for long-form ASR on LibriSpeech

• dreamerv3-robotic-control - DreamerV3 (29.5M params) for robotic control with RSSM world model, categorical latents (32x32), straight-through actor gradients, trained entirely in imagination on DMControl walker-walk
• action-conditioned-video-worldmodel - CogVideoX-5B fine-tuned with LoRA (rank 16) and novel action cross-attention adapter (95.2M trainable) for action-conditioned video prediction
• llm-planning-with-world-models - LLM planning with learned world models: Claude proposes action plans, DreamerV3 RSSM simulates outcomes, value head scores trajectories
• adaptive-inference-router-with-cascade-serving - Multi-objective RL routing across quantized/pruned model cascades
• adaptive-model-routing-serving-optimizer - Contextual bandit routing across FP32/FP16/INT8/pruned variants with 84.7% memory and 93.8% cost reduction
• adaptive-model-serving-optimizer - UCB bandit-based model serving with latency/accuracy/cost optimization
• adaptive-traffic-signal-control-via-hierarchical-multi-agent-rl - Hierarchical multi-agent RL for traffic signal control
• adversarial-curriculum-offline-rl-with-uncertainty-weighted-conservative-learning - Adversarial curriculum offline RL with conservative learning
• offline-robotic-manipulation-curriculum - Offline RL with curriculum learning for robotic manipulation
• temporal-distribution-shift-detector-with-adaptive-ensemble-reweighting - Bayesian online ensemble reweighting for distribution shift detection

A coordinated research program (AMPLE-GNC) on safe learned spacecraft GNC, each repo pairs a manuscript with full replication code, committed evidence, and a reproducibility manifest.

• paper-A-reliability-asymmetric - Reliability-asymmetric spacecraft GNC stack: a learned fault-adaptive controller and a language-model commander, bounded by a Kind 2-verified runtime shield, with end-to-end reproducible evidence
• paper-B-conformal-certificates - Conformal recovery-deadline certificates for runtime assurance of online-adapting controllers; distribution-free finite-sample coverage, demonstrated on a 6-DOF spacecraft and an inverted pendulum
• paper-C-ftc-controller-study - Honest settled-gate benchmark of learned vs classical spacecraft fault-tolerant control; a disturbance observer takes the additive-bias fault class from 0% to 59.4% with no SIGN/GAIN regression
• paper-D-assurance-transfer - "Assurance transfers, skill does not": re-certification cost of a verified runtime-assurance layer across spacecraft regimes (conformal certificate at 0 LOC changed, Kind 2 shield at 96% proof-form reuse)
• paper-E-flight-hil - Flight-class processor-in-the-loop validation on a Raspberry Pi 5 (24.5 µs control cycle, ~4000x under the 10 Hz budget; verified shield at 0.02% overhead) plus a high-fidelity fault campaign
• paper-F-settled-gate-benchmark - The settled-gate spacecraft fault-tolerant-control benchmark: a reproducible evaluation protocol and leaderboard for comparable learned and classical FTC
• paper-G-regime-ladder - Measuring skill vs assurance transfer cost across a six-rung spacecraft mission profile: assurance is regime-portable (~0 LOC), while skill is rebuilt only at actuator-family boundaries

Frameworks: PyTorch, HuggingFace Transformers, PyTorch Geometric, DGL, scikit-learn, XGBoost, LightGBM, FAISS, N-BEATS, PatchTST, TFT

Techniques: LoRA/PEFT, GNNs, contrastive learning, curriculum learning, uncertainty quantification, multi-task learning, DPO alignment, early-exit inference, spectral graph networks, GRPO reinforcement learning, knowledge distillation, Matryoshka representations, 3D Gaussian Splatting, rectified flow matching, Mamba SSMs, RF-DETR, speaker diarization, RNN-T

Infrastructure: MLflow, Docker, NVIDIA GB10 (128GB unified), NVIDIA RTX 4090, NVIDIA RTX 3090 (dual GPU), AMD Threadripper 3960X