Thanosan Prathifkumar

I study computer science at Caltech.

Currently, I do research on bayesian inference at Dartmouth College, computational imaging at the University of Toronto, and AI agents at the University of Waterloo.

Previously, I worked at RBC and Triage.

emaillinkedingithubresume

Work

Experience

  1. Software Engineering Intern · Royal Bank of Canada

    Jul 2026 – Aug 2026

    Built an LLM router that scores 30+ approved models on cost, latency, and accuracy — projected to save $792K/yr in inference costs.

  2. Founding Engineer · Constellation

    Jun 2025 – Sep 2025

    Built a founder registry from scratch: a canonical Postgres/Supabase schema, a magic-link profile editor, and an internal admin tool.

  3. Software Engineering Intern · Royal Bank of Canada

    Jul 2025 – Aug 2025

    Cut ATM restart time from 15 minutes to 10 seconds across 40+ servers with a Python and Ansible automation pipeline.

  4. Machine Learning Engineer Intern · Triage

    Apr 2024 – Dec 2024

    Built a medical-document RAG pipeline with LangChain and ChromaDB, grounding a clinical LLM in cited literature.

Build

Projects

  • Apr – May 2026

    GraphPlace: GNN + RL Macroplacement

    Beat the simulated-annealing baseline by 23.7% average proxy cost across all 17 IBM ICCAD04 benchmarks in the Hudson River Trading × Partcl macroplacement challenge, training a GNN policy that nudges macros out of congested regions.

    PyTorch · CUDA · C++ · Docker

  • Aug 2024 – Jun 2026

    Astraeus

    A radio-interferometry simulator computing UV coverage, dirty beams, and CLEAN reconstructions for ground and satellite arrays, with a U-Net reconstruction mode (SSIM 0.617).

    JAX · CUDA · ehtim

Writing

Writings

  • Publication· December 22, 2025
    Does SWE-Bench-Verified Test Agent Ability or Model Memory?

    University of Waterloo · First Author, with Noble Saji Mathews and Meiyappan Nagappan

    500+ experiments via the Anthropic API asking whether SWE-Bench-Verified scores reflect real coding-agent ability or memorized training data.

    Finds substantial overlap between benchmark tasks and model pretraining corpora, which complicates using the benchmark to compare agent capability.

  • Abstract· April 3, 2026
    Quantifying Uncertainty in Virtual Spatial Transcriptomics Using Bayesian Neural Networks

    Dartmouth College · AACR Annual Meeting 2026, San Diego (Abstract 4189)

    A Bayesian neural network that predicts gene expression from H&E tissue images and separates epistemic from aleatoric uncertainty across 991 genes and ~290k tissue spots.

    Metabolic and cell-cycle pathways turn out to be the most reliably inferred from morphology alone.