cv

B.Tech

Indian Institute of Technology Delhi

Electrical Engineering

• CGPA: 8.51/10 (~3.4/4.0 US scale)
• Probability & stochastic processes
• Digital image processing
• Machine intelligence & learning
• Data structures & algorithm
• Graph theory
• Computer architecture

Research Assistant

CodeMiner (Open Source)

Contributing to an open-source code retrieval system for AI coding agents, targeting VLDB.

• Built C++ indexing backend and inference optimizations for embedding and reranker serving.
• Improved token efficiency and reduced end-to-end retrieval latency for AI coding agents.

ML/AI Speech Engineer

Samsung HQ

Working on multilingual, low-latency ASR models and scalable C++ cloud infrastructure for Bixby.

• Fine-tuned a Qwen MoE streaming ASR model using LoRA adapters, investigating WER-latency tradeoffs under streaming constraints, achieving 8% WER on multilingual transcription.
• Developed a multilingual ASR model (31 languages) optimized for real-time streaming using quantization, chunking, K-V caching, VAD, and align-attention, achieving ≈ 45% latency reduction from baseline.
• Optimized ASR reliability under high-load conditions via advanced GPU synchronization techniques, achieving high hardware utilization and a 2% improvement in recognition accuracy during peak traffic.
• Implemented scalable C++ model handling modules within a multi-user Bixby Cloud environment for inference-optimized Fast Conformer-based ASR models (encoder-decoder and LSTM architectures).
• Designed an end-to-end robust data collection pipeline for low-latency ASR model training.

NLP Intern

Samsung HQ

Developed phonetic similarity metrics and database modules for speech recognition.

• Developed Name Entity Replacer for Personal Database module with tolerance toward mispronounced names.
• Proposed novel phoneme similarity metric based on manner of articulation paired with Metaphone and Soundex algorithm.
• Achieved 96.3% and 98.3% accuracy for spoken name matching, reducing recognition errors by ≈ 55%.

A Simple Model for Panic Driven Motion in a Refugee Camp

Journal of Human Behavior in the Social Environment (JHBSE)

Simulated crowd dynamics via numerical mollification and RK4; optimized obstacle configurations reducing fatalities by 70%.

LLM Scaling & Systems Optimization (JAX/TPU)

Engineered distributed Transformer components in JAX with TP, PP, and FSDP to overcome auto-compiler scaling bottlenecks.

• Hand-coded a Mixture of Experts (MoE) layer via jax.shard_map with explicit All-to-All token routing, completely eliminating native AllGather memory explosions.
• Leveraged XLA trace profiling to hide interconnect latency behind compute, achieving a 43x MoE speedup (73ms to 1.7ms) on TPU v5e-8.
• Workflow notebook (Kaggle)

Vision Semantic Communication

Developed a generative AI-based semantic communication pipeline for image communication targeting 6G, guided by Prof. Brejesh Lall, IIT Delhi.

• Proposed a guided Stable Diffusion + ControlNet approach using image captions, edge maps, and scaled-down latent representations.
• Achieved a 0.31 LPIPS score on VGG net in a noisy communication channel, a 40% improvement over the previous best.

Panic-Driven Crowd Dynamics

Developed strategies to reduce death count in dense crowd panic situations through simulation.

• Reduced death rate by 70%
• Applied numerical mollification and optimized obstacle shapes

DSCoin - Cryptocurrency

Built cryptocurrency system using SHA256 hashing and implemented mechanism to validate transactions.

• Proof-of-work Blockchain in Java
• Tree-structured Blockchain & broadcasting

JEE Advanced Rank 319

National Testing Agency

Secured All India Rank 319 among 150,000+ students.

Summer Undergrad Research Award (SURA)

Industrial R&D Unit, IIT Delhi

Conferred for excellent research.

JEE Main Rank 1383

National Testing Agency

Secured All India Rank 1383 among 1 Million+ students.