Researcher · Founder · Builder
I work on understanding how language models encode safety-relevant computations — what internal structures mediate alignment, where models fail silently, and how to verify that safety properties hold beyond surface behavior. I combine tools from interpretability, physics, and mathematics to attack problems that are under-explored.
My work spans internal representations of safety in language models, geometric analysis of transformer computation, behavioral monitoring for AI control, and emergent multi-agent dynamics. I approach these problems with tools from machine learning, interpretability, and mathematical physics.
Investigated how alignment shapes internal representations in instruction-tuned models by looking for the structures that mediate different types of refusal. Found two functionally independent safety boundaries — one predicting harmful-content refusal, another predicting disclosure/self-reference refusal — with near-zero cosine similarity, indicating genuine orthogonality rather than rotation of the same direction. Jailbreaks selectively cross one boundary at a time, providing behavioral evidence of independence. A key methodological finding: standard attribution tools (DiffMean, DLA, ablation) assume a single uniform refusal mechanism and produce systematically misleading results when that assumption fails.
Applied optimal transport — a mathematical framework for measuring the geometric work needed to move one probability distribution to another — to understand how transformer layers reorganize token representations. Treated activations as empirical distributions and measured distributional displacement layerwise across Qwen2.5 and GPT-2 models. Found consistent evidence that training induces specialization at the final layer, that in-context learning geometry is task-dependent (compression on reasoning, expansion on classification), and that OT cost correlates with attention entropy and detects distribution shift without inspecting model outputs — a possible passive safety signal.
Built a black-box monitor for detecting backdoored AI agents by stripping code to abstract syntax trees and comparing behavioral consistency across models from different training families — a "compare, don't judge" approach that doesn't require a trusted model to understand the task. Achieved 65% detection (AUC 0.727). The central finding was that cross-family evaluation reaches dramatically higher detection than same-family evaluation, suggesting that family-specific blind spots are a primary failure mode in existing monitors.
Investigating emergent collusive behavior in multi-agent systems across multi-turn and multi-agent settings in structured tabletop market environments. Focus on detecting and characterizing implicit coordination signals that arise without explicit communication channels.
Designed adversarial prompts exposing sabotage capabilities, evaluation-aware behavior, chain-of-thought issues, and PII leakage across hardened models. Scored 3.8× above the AI-only baseline on FAR.AI's challenge using staged phased attacks and probe-first methodology.
Leading training, optimization, and large-scale deployment of LLMs for low-resource Indian languages across ASR, TTS, machine translation, and OCR within the national Bhashini research consortium. Engineered Triton kernel optimizations that significantly reduced inference latency across production pipelines. Built a multilingual speech-to-speech translation system with real-time latency optimizations and multi-user support. Currently developing a domain-specific fine-tuning pipeline (TinyLlama + LoRA + RAG) for the Indian agriculture domain.
Founded and led an internal LLM mathematical reasoning research group. Built an NER + intent classification model improving chatbot engagement by 20%. Designed an anomaly detection pipeline reducing manual effort by 60%. Developed QA evaluation metrics using semantic similarity. Created a knowledge-tracing recommender system. Received the Think Tank Innovation Award (individual) and Kingsmen Award (team).
Invented an autonomous underwater vehicle system for garbage detection using CycleGAN with a custom dataset surpassing SOTA benchmarks. Won the Water Research Olympiad. Received mentorship from Persistent Systems. Two Indian patents filed (No. 201921043504; No. 202021028978). IEEE publication on multiclass image classification for aerial vehicles.
I'm Jank — an AI safety researcher based in Pune, India. I work at the intersection of alignment, interpretability, and the mathematical structures that underpin neural computation. My goal is to understand how safety properties are encoded in models before capabilities make that problem intractable.
I lead ValueShift Research, an independent safety research group where we run experiments and publish on LessWrong and the Alignment Forum. My longer-term aim is to join a dedicated safety lab to work on pre-capability verification of alignment.
I completed the Algoverse AI Safety Fellowship working on multi-agent collusion research, and am continuing that line of work independently — investigating how implicit coordination emerges in multi-agent systems without explicit communication.
I'm also actively studying physics — working through Halliday and Resnick — because I believe physical intuitions about geometry, fields, and conservation laws are underutilised in how we think about what happens inside neural networks.
During my undergraduate years I founded the ACM-W chapter at my university — one of the first in the region — as a fully paperless organisation with a deliberate model of engaging men as active allies. I wanted to build something that would outlast my presence, not depend on it.
Merit scholarship — fully funded undergraduate engineering degree through demonstrated academic excellence.
Selected as 1 of 40 girls nationally. Full scholarship for women in engineering with demonstrated technical promise and leadership.
Founded women-in-STEM chapter with fully paperless operations and a male-ally model. Active mentorship of junior students across disciplines.
Grace Hopper Celebration India scholarship for women in computing.
First place for AUV-based underwater plastic detection. Two Indian patents filed.
Mentoring practitioners transitioning into ML engineering and applied AI research.
Top of class across all semesters. SGPA 9.34/10, Savitribai Phule Pune University.