Neha S — AI Safety Researcher

Experience

DRDO DYSL-AI

Aug 2025 – April 2026 · Bengaluru, KA

Architected a multilingual evaluation benchmark spanning 5+ languages, surfacing a 15-point safety gap between high- and low-resource languages.
Curated a specialized evaluation dataset probing model behavior in sensitive domains, lifting refusal consistency by 20%.
Quantified cross-lingual alignment disparities to shape safety-guardrail refinements for regional-language LLM deployments.

IIITDM

May 2023 – June 2023 · Kancheepuram, TN

Optimized Intrusion Detection Systems for HPC, reducing false-positive alerts by 20% through vulnerability mapping.
Authored a bug-bounty methodology report identifying 3 critical infrastructure gaps.

M.Tech Thesis · DRDO-DIAT × DRDO DYSL-AI

Designed a six-stage model-agnostic safety pipeline wrapping 6 open-source LLMs without modifying weights, deployed on a single RTX 4090.
Built an 11-category, 6,000-sample harm-intent dataset from 6 public benchmarks + 360 custom jailbreak prompts.
Fine-tuned RoBERTa-base classifier reaching 95.6% accuracy (0.958 macro F1) — a 34-point gain over keyword baseline.
Cut jailbreak success rate from 23.3% → 3.3%; confirmed on AdvBench, HarmBench, XSTest.
Shipped a production FastAPI service with 5 REST endpoints, audit logging, and a React monitoring dashboard.

B.Tech Final Year · AMC Engineering College

Full-stack virtual try-on app combining live pose tracking, GAN-based clothing synthesis, and visual recommendations.
Used OpenCV + OpenPose for body/pose detection; adapted CP-VTON/ACGPN-style GAN for photorealistic garment synthesis at 256×192.
Recommendation engine using ResNet50 embeddings + KNN (k=6) for visually similar product retrieval.
Validated across 6 testing levels — unit, functional, UI, compatibility, navigation, integration.

Design and Development of Autonomous Driving Car Using Nvidia Jetson Nano Developer Kit

Lecture Notes in Networks and Systems, vol 1116 · Springer, Singapore

Co-engineered an autonomous JetRacer on Jetson Nano with PyTorch, OpenCV, TensorRT and SSD-Mobilenet/SSD-Inception networks (MS-COCO, 90 classes).
Accelerated inference via torch2trt; improved obstacle-recovery accuracy through iterative fine-tuning on track datasets.

PythonC++SQLJavaScript

PyTorchTensorFlowHuggingFaceScikit-learnOpenCVLoRA Fine-tuning

RLHFDPOPPOJailbreak DetectionAdversarial RobustnessReward Modelling

FastAPIPostgreSQLMongoDBCassandraHadoop

GitDockerLinuxAWSTensorRTJetson Nano

Aug 2024 – May 2026

Defence Institute of Advanced Technology (DRDO-DIAT)

Focus: ML, LLM Safety, Algorithms · CGPA: 8.0/10

Aug 2020 – May 2024

AMC Engineering College (VTU)

Focus: Neural Networks, Computer Vision · CGPA: 8.4/10