Curriculum Vitae

B.Tech in Computer Science and Engineering (Hons.) · CGPA 9.6 / 10 · Top 10 of the graduating class

With Prof. Ross Shachter

• Designed experiments testing whether LLMs can execute decision analysis, and discovered that response-field order alone changes the answer: $60, $70, $82.35 and $67.50 under answer-first elicitation versus a stable $76.00 under reasoning-first, a measurement artifact that affects how LLM reasoning is evaluated.
• Built a code-generation setup in which models solved decision problems exactly (128/128) where in-context reasoning failed, and showed that error feedback lifts GPT-4o from 10.2 to 23.0 of 29 tests but never to perfection.

With Dr. Xiangrui (Taylor) Zeng

• Lead a team generating patient-specific 3D brain MRI volumes without paired data using diffusion models in PyTorch. Set the research direction by hypothesising that gradient guidance would fail where adapter conditioning succeeds, and designed an oracle experiment to test it.
• Traced residual blur to conditional-mean averaging over plausible anatomies, localised it to high-frequency wavelet subbands, and designed a wavelet-domain loss that improved sharpness; built the ADNI data pipeline and nnUNet-based segmentation evaluation pipeline, generating all second-phase results.

• MIRA (CVPR 2026 workshop, advisor Yunbei Zhang): evaluated 11 frontier VLMs on a benchmark of 713 real images and 3,529 question-answer pairs across 10 scenarios, finding they scored 39–48% against 94% for humans, and that multi-turn dialogue degraded most models.
• SR-LoRA / CLAC (ICLR 2027 submission, with Xingjian Li): built and ran the 480-compositions-per-method GPU benchmark on Stable Diffusion v1.5 (rank-4 LoRAs, DPM-Solver++, CLIP ViT-L/14) and the τ/α/δ and N=3 ablations; the ten-seed run showed the text-alignment gain was 4.3%, not the 6.8% seen with three seeds. CLAC raises identity consistency 0.839→0.915, 7.6 points over the strongest baseline.
• TRACE and APISHIFT: contributed to verifiable reinforcement learning environments, including environment design, verification logic and reward mechanisms that score agent actions against ground truth rather than relying on a judge model.

With Dr. K. Ashesh

• First-authored an IEEE CONIT 2024 paper applying outlier analysis to credit card fraud detection.

Pune, India

• Architected and built an 8-microservice platform (Java 21, Spring Boot, PostgreSQL, AWS) centralising vulnerability management, FinOps, database operations and infrastructure automation, adopted by 100+ applications and 50+ engineering teams.
• Built LangGraph agents with MCP tool integrations that run security scans, detect CVEs, create Jira tickets and generate reports automatically, replacing manual triage across 50+ teams; recognised with the Director's Recognition Award for improving automation reliability and reducing user-prone errors.
• Built automation covering 15,000+ services and 4,000+ applications for container remediation, vulnerability detection and validation pipelines, cutting remediation time from days to under one hour.
• Designed a hybrid SES–Keycloak authorisation model with dynamic role provisioning, giving 50+ teams self-service onboarding.
• Developed an AWS FinOps module for multi-account cost analytics, anomaly detection, forecasting and savings recommendations; automated GitLab migration and CI/CD onboarding for 100+ repositories.

Pune, India

• Built a full-stack cache monitoring platform (Spring Boot, Thymeleaf, Apache Ignite) tracking SDS and ESM cache data across distributed systems, with anomaly detection and automated email alerts on cache fluctuations; improved data retrieval speed by 30%.

Hyderabad, India

• Built a backend platform for NGO operations with CRUD APIs for donor management, beneficiary tracking and fund distribution, improving SQL query performance by 40%.