I'm a Data Science undergraduate at IIT Madras interested in understanding how machine-learning systems behave — especially when they encounter noisy data, distribution shifts, unreliable retrieved context, or adversarial inputs.

My work currently spans:

• LLM & agent evaluation — instruction–data separation, tool-use reliability, critique agents, and adversarial evaluation
• Temporal ML — sequence modeling with TCNs, LSTMs, and MS-TCNs
• Retrieval systems — RAG, embeddings, vector databases, and evidence-grounded generation
• Reproducible experimentation — controlled evaluations, failure analysis, structured logging, and model comparison
• Open models — experimenting with open-weight models and transparent evaluation pipelines

I like building experiments that make model failures easier to reproduce, measure, and understand.

                 ┌──────────────────────┐
                 │   Model Behaviour    │
                 └──────────┬───────────┘
                            │
              ┌─────────────┼─────────────┐
              ▼             ▼             ▼
        Agent Reliability  Retrieval   Temporal ML
              │             │             │
              ▼             ▼             ▼
       Adversarial Eval   Grounding   OOD Evaluation
              └─────────────┬─────────────┘
                            ▼
                 Reproducible Experiments

I'm particularly interested in questions like:

• When do tool-using agents treat untrusted data as instructions?
• Can critique or verification stages reliably reduce unsupported generation?
• How should model failures be measured beyond aggregate accuracy?
• How do sequence models behave when demonstrations differ from their training distribution?
• How can evaluations be made easier for other researchers to inspect and reproduce?

Mar 2026 – May 2026

Worked on temporal modeling of human-manipulation demonstrations using TCN, LSTM, and MS-TCN architectures.

The experimental pipeline covered:

video data → preprocessing → kinematic features → temporal models → evaluation → error analysis

The dataset contained approximately 9,000 real-world videos across six action classes, including evaluation on out-of-distribution demonstrations.

flowchart LR
    A[Question] --> B[Controlled Experiment]
    B --> C[Structured Evaluation]
    C --> D[Failure Analysis]
    D --> E[Iterate]
    E --> B
    D --> F[Document Results]

I prefer experiments where the evaluation setup, assumptions, metrics, and failure cases are visible, rather than treating a single benchmark number as the entire result.

I'm currently spending most of my project time on:

LLM evaluation → adversarial behavior, critique systems, grounding
Agent reliability → instruction–data separation and tool-use failures
Open models → experiments that can be inspected and reproduced
Retrieval → evidence-grounded generation and failure-aware pipelines
Temporal ML → sequence modeling and generalization

I'm interested in conversations around ML research, open models, evaluation, agent reliability, retrieval systems, and reproducible experimentation.