cv

Member of Technical Staff

Parallel Web Systems Inc.

AI agents for deep web research

• Built Parallel's deep web research agents processing >1M queries/day - achieved #1 on BrowseComp (58% vs 53% runner-up), becoming key highlight of company's public launch
• Architected browser automation translating DOM to LLM-readable format and converting agent decisions to Playwright actions, enabling form navigation and CAPTCHA handling via proxies
• Built citation excerpting and answer confidence modules that identify supporting URLs/excerpts and calibrate answer reliability - achieved >66% accuracy on high-confidence answers vs <33% on low-confidence, enabling users to trust agent outputs and powering workflow decision logic
• Building self-improving LLM framework to gradually replace commercial APIs with finetuned open-source models using LLM-as-judge validation - targeting 40% cost and latency reduction across all Parallel agents
• Built vision LLM pipeline converting screenshots to markdown, enabling agents to extract structured data from tables, graphs, and visual content
• Created domain crawlers templatizing agent actions (pagination, infinite scroll) for exhaustive entity extraction - core technology powering FindAll product ('SQL over the web')

Internship + Full-time Machine Learning Engineer

Granica

Language models for data anonymization

• Led ML team building Granica's PII de-identification product - achieved SOTA performance processing 10TB/day for Fortune 500 healthcare and insurance clients
• Redesigned tabular PII detection from regex to finetuned LLM embeddings, doubling accuracy from 40% to 80%
• Innovated teacher-TA-student distillation framework for free-text PII detection - solved BERT-GPT architecture mismatch by training intermediate TA model before distillation, delivering 1.5x recall and 3x precision
• Optimized production BERT inference 3x through vLLM-inspired server-side batching, reducing costs at 10TB/day scale

Software Engineer

Samsung Electronics

Built ML-powered tools for device diagnostics and issue resolution

• Built ML-powered issue routing system used by 200+ Samsung engineers - reduced issue resolution time from 1 week to 1 day using language models and heuristics
• Deployed router handling dozens of user-reported device issues daily, automatically directing them to correct engineering teams
• Developed log analysis tool trained on historical issue data to highlight potential problem areas in device logs for debugging

Master of Science

Stanford University

Computer Science (AI concentration)

Bachelor of Technology

Indian Institute of Technology, Bombay

Computer Science and Engineering

GPA: 9.86/10.0 (Rank 1, University Silver Medalist)

MC-MED, multimodal clinical monitoring in the emergency department

Scientific Data

Kansal, A., Chen, E., Jin, B.T., Rajpurkar, P., & Kim, D.A. (2025). MC-MED, multimodal clinical monitoring in the emergency department. Scientific Data, 12(1), 1094.

Multimodal clinical benchmark for emergency care (MC-BEC)

NeurIPS

Chen, E., Kansal, A., Chen, J., Jin, B.T., Reisler, J., Kim, D.E., & Rajpurkar, P. (2023). Multimodal clinical benchmark for emergency care: A comprehensive benchmark for evaluating foundation models in emergency medicine. Advances in Neural Information Processing Systems, 36, 45794-45811.

Error-driven fixed-budget ASR personalization for accented speakers

ICASSP

Awasthi, A., Kansal, A., Sarawagi, S., & Jyothi, P. (2021). Error-driven fixed-budget ASR personalization for accented speakers. ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing.

Multi-tiered Approach to Debiasing Language Models

Developed novel debiasing techniques for BERT targeting gender, race, and religion biases using CDA and SentenceDebias variants

• Achieved highest ICAT scores with SDB-Equalized method that preserves information by equalizing embeddings across bias subspace instead of removing components
• Discovered high cosine similarity between gender/race/religion bias subspaces, showing biases share common representation - unified model outperformed type-specific models
• Implemented 6 debiasing variants including CDA+GPT (using GPT-4 for stereotype-antistereotype pair generation) and SDB-Unified (multi-bias simultaneous correction)
• Evaluated on StereoSet and CrowS-Pairs benchmarks - SDB methods consistently outperformed CDA by 5-10% on bias metrics while maintaining language modeling performance

Fraudulent Activity Recognition in Graph Networks

Built GNN-based fraud detection system for financial networks and e-commerce using node classification on large-scale graphs

• Achieved 75.94% AUROC on DGraph-Fin (3.7M nodes, 78:1 class imbalance) and 95.55% AUROC on Amazon Reviews using graph-based fraud detection
• Implemented 5 GNN architectures from scratch including ChebNet (spectral graph convolutions) and PC-GNN (imbalanced learning) using PyTorch Geometric
• Developed heterogeneous r-GNN models (r-GCN, r-GraphSAGE, r-ChebNet) to model 3 relationship types in Amazon product review networks
• ChebNet achieved near-best performance with 7× fewer parameters through multi-hop neighborhood aggregation in single layer using Chebyshev polynomials

3D Pose Analysis: Predict, Match & Search

Built end-to-end system for 3D pose estimation and matching using state-of-the-art DiffPose and POEM models

• Integrated DiffPose (diffusion-based 3D pose estimator) with POEM (pose embedding model) using transfer learning - achieved superior performance over baseline ICP algorithm
• Developed interjoint distance embedding method (120-dim) for direct 3D pose comparison, outperforming traditional point cloud registration approaches
• Curated open-source dataset of 408 images across 8 pose classes from Google Image API with rigorous filtering for pose consistency and environmental variation
• Achieved robust pose matching with scale/translation/rotation invariance, successfully handling occlusions like 'hands in pocket' vs 'hands on hips'

GAN-Inspired LLM Framework for Spear Phishing Detection

Built adversarial LLM system for automated phishing email generation and detection using self-improving prompts

• Developed GAN-inspired framework with GPT-4 discriminator and Llama2 generator achieving 77% fooling rate against human evaluators
• Designed self-prompting mechanism that autonomously generates and refines phishing email prompts through adversarial training, eliminating manual prompt engineering
• Introduced novel evaluation metrics adapted from RAG: Groundedness, Phishing Score, Guideline Correlation (r=0.693), and Accuracy
• Trained on 100 real emails with dynamic guideline compaction algorithm to prevent redundancy and maintain efficiency across epochs

X-ray-to-Report Conversational AI

Built a multi-modal vision-language model for automated radiology report generation and interactive patient Q&A from chest X-rays

• Developed conversational AI using LLaVa (CLIP + LLaMa-7B) fine-tuned on 356K+ MIMIC-CXR image-report pairs - achieved 40% improvement in BLEU score for report generation
• Created novel QA dataset by processing radiology reports through ChatGPT API, enabling interactive multi-turn conversations about X-ray diagnoses
• Implemented parameter-efficient fine-tuning using LoRA and 8-bit quantization to train 7B parameter model on consumer GPUs
• Achieved accurate diagnosis of conditions like pleural effusion and lung masses through qualitative evaluation, outperforming zero-shot baseline model

MLETA: Meta Learning for Efficient Test-time Adaptation

Developed a novel MAML-based Test-Time Adaptation approach for domain adaptation in image classification

• Achieved 3.7% improvement in mean classification error over baseline TTT method on CIFAR-10-C benchmark with 15 corruption types
• Designed meta-learning framework using MAML to explicitly optimize feature extractor and classification head synchronization during test-time adaptation
• Delivered 3× faster inference (40.7ms vs 118.7ms) while outperforming baseline that used 3× more adaptation steps
• Implemented and evaluated multiple self-supervision strategies (rotation prediction, flip prediction, entropy minimization) for unsupervised domain adaptation

Institute Silver Medal

IIT Bombay

Awarded to the top-ranked student in the Computer Science and Engineering department

Charpak Lab Scholarship

French Ministry of External Affairs

Highly selective research scholarship for Indian students to conduct research in France (~30-50 awardees annually)

Institute Academic Prize

IIT Bombay

Top 1% academic performance

JEE Advanced - All India Rank 15

Joint Entrance Examination

Ranked 15th among 200,000+ candidates in India's premier engineering entrance examination

JEE Main - All India Rank 21

Joint Entrance Examination

Ranked 21st among 1.2 million candidates in national engineering entrance examination

BITSAT All India Topper

Birla Institute of Technology and Science

Secured 1st rank with a score of 454/450 among 190,000+ candidates

National Science Olympiads

HBCSE (Homi Bhabha Centre for Science Education)

Chemistry Olympiad: Gold Medal (top 35 out of 39,400 candidates), qualified for OCSC. Physics Olympiad: Top 1% nationally, qualified for INPhO. Astronomy Olympiad: Top 1% nationally, top 34 students, qualified for OCSC

KVPY Scholar

Indian Institute of Science

Kishore Vaigyanik Protsahan Yojana - National science scholarship (~1% selection rate, ~1,000 from 100,000+ applicants)

NTSE Scholar

National Council of Educational Research and Training

National Talent Search Examination - Top 0.1% nationally (~1,000 from 500,000+ applicants)