aritro shome avatar — illustrated portrait with a cat on the head

aritro shome

(or sortira)

student researcher with interests in multimodal/multilingual interpretability, alignment & safety. also: reasoning models, RL, synthetic data, making smaller models better, for democratized ai.

currently research intern at AI4Bharat (visual reasoning, multimodal multilingual interp). previously ai engineer intern at Sarvam AI — built the eval framework for an NVIDIA lipsync model + modularized the dubbing pipeline (video).

i like cracking models open to figure out why they fail, analogous to a neurosurgeon of ai models.

writing silicognition.is-a.dev my technical blog on solo research projects, encounters with ai.

experience

  1. [mar 2026 - present] research intern @ AI4Bharat — circuit tracing for reasoning problems posed in various modalities across model sizes, languages for various SoTA open models.
  2. [aug 2025 - dec 2025] ai engineer intern @ Sarvam AI — eval framework for NVIDIA lipsync; detected and fixed 500% file size increase in lip-sync model output; modularized dubbing pipeline; a/b tests on prompt structures

solo research projects

  1. forgeformer — hand-designed 2-layer transformer that adds two-digit numbers; no gradient descent. demonstrated transformer internals in a model designed with interpretable weights from ground-up.
    [writeup] · [demo]
  2. seven deadly sins of gemma — steering vectors per sin in gemma-2-2b; verified the vectors via steering examples and activation patch analysis. [writeup]
  3. some of the works done with other orgs are not publishable.

education

  1. [2024 — 2028] b.tech, information technology @ IIEST ShibpurCGPA 9.15
  2. [2010 — 2024] m. p. birla foundation higher secondary school — ISC 95.3% / ICSE 98.3%
  3. INSPIRE scholar — Govt. of India (awarded to top 1% in the higher-secondary board qualifying exam)