Independent research program
Independent Research Scientist
My work spans five connected directions: foundation models, computer vision, uncertainty quantification and risk control, computer graphics, and intelligence.
Research agenda
Foundation models
What representational structure makes broad transfer possible? This line tests whether recurrent, compositional representations are a prerequisite for foundation-model behavior beyond language.
View public workComputer vision
How can visual evidence support inspectable decisions and structured world models? This line spans finite-sample auditing, symbolic 4D scene representations and persistent motion models for video and LiDAR.
View public workUncertainty quantification and risk control
Finite-sample calibration, distribution shift, expert disagreement and sequential decision workflows across medical, social and incident-analysis settings.
View public workComputer graphics
Structured representations for static and dynamic 3D scenes, including executable scene grammars, Gaussian scene experiments and procedural motion.
View public workIntelligence
Two connected programs: a general theory of how natural intelligence constructs and recursively reuses abstractions, and a focused investigation of whether abstraction can reduce computational complexity.
Explore the research directionResearch projects and artifacts
Current projects and public repositories connect the agenda to concrete methods, experiments, software and explicit limitations.
View the complete GitHub profileFoundation models
What makes a representation foundation-ready, so that learning can accumulate across contexts and familiar units can support genuinely new combinations?
MotifShift
Controlled experiments test cross-context unit reuse, distributional stability, sequence recombination and transfer across trajectories, gridworld replay and physical fields.
View repository Visual representationsReusable Vision Representations
Measures selectivity, cross-context portability and held-out compositional transfer, including where specialization through depth begins to reduce reuse.
View repositoryComputer vision
Auditing visual evidence and learning structured representations of scenes and motion from images, video and LiDAR.
Finite-Sample Vision Auditing
Exact decision and influence certificates, contextual contribution estimates and integrity-checked experiments for studying spatial decision stability and reserve evidence.
View repositoryBeyond the Radiance
Spatio-temporal scene grammars that turn visual evidence into compact, editable programs of structure, motion and interaction.
Learning 3D Tapes from Video
A symbolic, compositional representation that joins spatial layout, object semantics and procedural motion for replay, editing and simulation.
Motion Tunnels
Probabilistic 4D motion representations designed to preserve object identity and continuity through sparse observations and occlusion.
Uncertainty quantification and risk control
When does a calibrated policy remain useful, how does it fail under changing conditions, and what evidence is needed to recalibrate it?
Shift-Aware Risk Control
Global, context-transfer and context-weighted risk-control methods with audits for simultaneous validity, transfer penalties and data leakage.
View repository Sequential workflowsConformal Pathways
Result-aware branching workflows that measure reference retention, selected-rule violations, abstention and diagnostic-test use separately.
View repository Consequence-aware evaluationMedical GenAI Beyond Factuality
Separates evidence support from potential clinical consequence and compares consequence-selective evaluation with uniform factuality filtering.
View repository EyeTrustAI research agendaClosed-Loop Conformal Risk Control
A broader program on how risk certificates transfer, govern action, receive feedback and adapt under changing deployment conditions.
View research agendaComputer graphics
How can scenes and motion be represented as inspectable, editable and reproducible structures rather than opaque outputs?
SceneGrammarKit
Deterministic split-and-repeat grammars for coarse semantic scene structure, with typed edits, bounded execution and headless rendering.
View repository Structured 3D Gaussian scenesGaussWeave
Reproducible tooling for structured 3D Gaussian experiments, including deterministic scenes and cameras, rendering adapters and artifact-integrity checks.
View repository Dynamic 3D representationKinematicWeave
Persistent spatial records and procedural motion with exact and error-bounded trajectory codecs, semantic events and matched-budget evaluation.
View repositoryActive research portfolio
Conformal risk control for deployed AI systems
This program asks how statistically valid risk guarantees can survive the conditions of real deployment: changing environments, adaptive policies, selective feedback, delayed labels and systems whose decisions alter the data they later observe.
Certificates across environments
How can a risk certificate remain meaningful when an AI system moves beyond the environment in which it was calibrated?
- Contextual selective CRC: transferring release-risk guarantees to a new deployment context.
- Meta-conformal risk transfer: certifying learned relationships across a population of environments.
- Certificate survival under drift: determining when validity decays and a certificate must expire.
- Adaptive risk contracts: allowing loss definitions, subgroups and evidence rules to change without invalidating guarantees.
Closed-loop AI control
How can risk be controlled when AI actions, human review and the resulting evidence form a feedback loop?
- Adaptive sequential conformal authorization: allocating risk across multi-step AI workflows.
- CRC under policy-dependent feedback: preserving guarantees under selective release, delayed labels and review bias.
- Safe active auditing: allocating human review to improve future autonomy while retaining risk guarantees.
- Performative CRC: studying risk guarantees when the controller changes the environment it controls.
Closed-Loop Conformal Risk Control
A unified theory combining context shift, adaptive policies, selective feedback, certificate expiration and recalibration for deployed AI systems.
Medical imaging for clinical deployment shift, autonomous systems and robotics for sequential control, and AI agents for multi-step language-enabled workflows.
Selected publications by research area
The selection below is organized around the first four areas of the research agenda. The complete indexed record is available on DBLP.
Foundation models
- Motion as Language: Towards a Situation–Motion Language for Spatio-Temporal Learning. NeurIPS Workshops, 2025.
- Ophthalmology as a Lens for Trustworthy GenAI in Europe. NeurIPS Workshops, 2025.
- CLEAR-Command. NAACL System Demonstrations, 2025.
- AEXL: Enhancing Path Prediction with Active Explainable Learning via Large Language Models. CHI Extended Abstracts, 2025.
Computer vision
- DeSPITE: Deep Skeleton–Pointcloud–IMU–Text Embeddings. ICCV, 2025.
- NeSyMoF: A Neuro-Symbolic Model for Motion Forecasting. IROS, 2024.
- LiOn-XA: Unsupervised Domain Adaptation via LiDAR-Only Cross-Modal Adversarial Training. IROS, 2024.
- Unsupervised 4D LiDAR Moving Object Segmentation. WACV, 2023.
- Improving Wearable-Based Activity Recognition Using Image Representations. Sensors, 2022.
- VIDENS: Vision-Based User Identification from Inertial Sensors. ISWC, 2021.
Uncertainty quantification and risk control
- SafePath: Conformal Prediction for Safe LLM-Based Autonomous Navigation. arXiv, 2025.
- Conformal Prediction for Semantically-Aware Autonomous Perception in Urban Environments. CoRL, 2024.
- AR-CP: Uncertainty-Aware Perception in Adverse Conditions. CVPR Workshops, 2024.
- Can You Handle the Truth? Investigating AR-Based Communication of Deep-Learning Uncertainty. ISMAR, 2023.
- Visualization of Machine Learning Uncertainty in AR-Based See-Through Applications. AIVR, 2022.
Computer graphics
- Whenever, Wherever: Orchestrating Crowd Simulations with Spatio-Temporal Spawn Dynamics. ICRA, 2025.
- PointCloudLab: 3D Point Cloud Annotation with Visual Aids and Immersion. ICRA, 2023.
- VR-Surv: A VR-Based Privacy Preserving Surveillance System. CHI Extended Abstracts, 2022.
Earlier work in software engineering, continuous experimentation and pervasive systems remains part of the broader publication record and will be integrated more fully in a later revision.