Navish Kumar

I study how learning systems change—and build tools that keep that change inspectable.

Machine-learning researcher and systems builder at the University of Basel. My work connects optimization, continual learning, mathematical structure, and evidence-grounded systems.

View research Get in touch
  • Continual learning
  • Optimization
  • Reliable intelligent systems

01 / Research

Current and published research

Work on the geometry of learning, constrained adaptation, graph structure, and decision systems. Status labels link claims to their public record.

  1. Under revision

    Experience Replay Through the Lens of Optimization

    Frames replay as constrained correction matching and makes residual mismatch observable when a memory buffer cannot realize the intended update.

    With Aurelien Lucchi

    OpenReview
  2. Under review

    Rank Feasibility in Continual PEFT

    Tests whether a low-rank adaptation space contains a task-wise correction before spending the optimization budget.

    With Aurelien Lucchi

    OpenReview
  3. Preprint

    Optimization Guarantees for Square-Root Natural-Gradient Variational Inference

    Convergence guarantees for variational Gaussian inference using a square-root covariance parameterization.

    With Thomas Möllenhoff, Mohammad Emtiyaz Khan, and Aurelien Lucchi

    arXiv
  4. Workshop paper

    Modelling Delivery Vehicles Across Urban Micro-Regions

    A spatial modelling pipeline for evaluating cargo-bike transition at the scale where operational conditions change.

    arXiv
  5. Published

    Normalized Laplacians for Gain Graphs

    A normalized spectral framework for complex unit gain graphs, including balance, interlacing, and structural characterizations.

    Journal
  6. Published

    Bounds for the Extremal Eigenvalues of Gain Laplacian Matrices

    Connects extremal spectral bounds to graph inconsistency, frustration, and structural repair cost.

    Journal
  7. Published

    Interaction Dynamics Between Hate and Counter Users on Twitter

    A paired-user dataset and empirical analysis of interaction, linguistic, and behavioural asymmetries.

    ACM

02 / Selected work

Systems and open source

A small set of maintained or research-defining projects. Experiments that do not sharpen this record stay out of the portfolio.

01

CasePath

Evidence-grounded claim handling that separates model interpretation from deterministic authority, provenance, and review.

  • Agent systems
  • Product engineering
  • Verification
Source Live demo

02

Green Last Mile

Open spatial modelling work for comparing delivery modes across heterogeneous urban micro-regions.

  • Applied ML
  • Urban systems
  • Spatial data
Source Paper

03

Hate and Counterspeech Dynamics

Code and data for studying paired hate and counter-speech accounts as an interaction system.

  • Social computing
  • Networks
  • Empirical research
Source Paper

04

Gain-Laplacian Spectral Bounds

Reproducible experiments accompanying work on extremal eigenvalues and structural inconsistency in gain graphs.

  • Spectral graphs
  • Optimization
  • Reproducibility
Source Paper

03 / Experience

University of Basel

Research grounded in explicit assumptions, observable diagnostics, and implementations that can be checked.

2022—Present

Basel, Switzerland

PhD Researcher

I work across continual learning, variational inference, and robust optimization, building evaluation pipelines that connect theoretical guarantees with operational constraints.

University profile

04 / Contact

Let’s discuss research, systems, or a difficult problem worth making precise.

navish.kumar@unibas.ch