Aritro Shome

AI4Bharat, IIT Madras

• Investigating mechanistic interpretability and circuit tracing of SoTA open models (Gemma, Qwen) on reasoning tasks posed across modalities, languages and model scales — focused on vision-language models over Indic inputs.
• Ran 40,000+ instrumented forward passes for activation capture on multi-billion-parameter models under a hard 24 GB VRAM ceiling, and built pre-experiment component-attribution scripts to isolate critical circuits before committing to full runs.
• Engineered a memory-efficient activation-patching framework: disabled fused and attention-kernel optimisations to expose per-layer activations, then managed the resulting non-linear VRAM growth to keep large models inside budget.

Sarvam AI — media & dubbing

• Built an automated evaluation framework for an NVIDIA lip-sync model; the quantitative pipeline caught and resolved a bug causing ~500% variance in output video file sizes.
• Redesigned the production dubbing pipeline — replaced brittle hard-coded prompt harnesses with modular YAML instruction blocks that let the model's chain-of-thought infer context, raising translation quality; picked the final configuration by A/B test.

• Hand-designed an interpretable 2-layer transformer that performs two-digit addition with manually-set weights and no gradient descent — building attention and MLP computation from first principles to show exactly what each component contributes.

• Replicated Anthropic's persona / emotion-vector methodology to extract a steering vector for each of the seven deadly sins in Gemma — on a single consumer GPU.
• Verified vector correctness through targeted steering and activation-patching experiments rather than eyeballing generations.

• Built a modular Python engine that generates alpona — Bengali folk art — through fractal-based layering, with an extensible OOP style-registration system.
• Published a 13,000-image synthetic dataset at 1024×1024 on Kaggle for generative-AI research.