Research & Experience
I work on the internals of machine-learning systems — interpretability, retrieval, and structure in relational data — across independent research and internships.
Working papers
Mechanistic Interpretability of Transformer Language Models
In preparationV. Sathveek · Independent research · 2025
Interpretability experiments on GPT-2 Small — activation caching, attention-pattern analysis, residual-stream decomposition, and induction-head detection — using activation patching as a causal intervention to identify which attention heads drive specific predictions.
Experience
Software Engineer Intern
May 2026 – PresentIntants Pvt Ltd · Remote
- Automated extraction, cleaning, validation, and transformation of vendor invoice data on a full-stack platform (Python/FastAPI, PostgreSQL, React), removing manual data-entry effort across the finance workflow.
- Designed a normalized PostgreSQL schema and wrote optimized SQL for invoice ingestion and line-item reconciliation, improving accuracy and throughput under concurrent load.
- Built role-based review and approval dashboards with real-time status tracking, giving the finance team clear visibility into invoice pipelines.
Data & Backend Developer Intern
Aug 2025 – Dec 2025National Institute of Technology, Andhra Pradesh · Tadepalligudem, Andhra Pradesh
- Built data-analytics pipelines in Python (pandas, NumPy) over user activity and performance data, and delivered interactive dashboards of topic-level trends that reduced student preparation time by ~30%.
- Designed a question prioritization and tagging system driven by user profiles and historical performance, powering personalized recommendations that improved practice efficiency by ~20%.
Research Intern
May 2025 – Aug 2025Indian Institute of Technology, Ropar · Ropar, Punjab
- Modeled large relational datasets as graphs to analyze relationships and dependencies, surfacing network patterns and structural insights not visible in raw tabular data.
- Built an interactive visualization interface with filtering and toggling for real-time exploration of relational data, improving insight discovery for researchers.