Sep 2023 – Sep 2024 · 1 yr · Bengaluru, Karnataka, India · Hybrid

• Client-Centric Foundation Model LLM Implementation
• Pioneered LLM use-case development, optimizing AI solutions for diverse client needs.
• Championed prompt engineering and RAG adherence, ensuring quality and industry compliance.

Jan 2023 – Sep 2023 · 8 mos · Bengaluru, Karnataka, India · On-site

• 1. Hand Keypoint Detection:
• Trained neural networks for hand keypoints in AR/VR using TOF, achieved 6ms DSP runtime.
• Leveraged trihorn-net for 2D heatmaps and 2.5D hand coordinates, improving MPJP error to 8 pixels.
• 2. Data Collection & Ground Truth Generation:
• Designed a 6-camera system for hand keypoints. Used PoseNet models for initial keypoints with 2D visualization corrections.
• Captured pinch gestures and self-occlusion, collecting 1 lakh data points.
• 3. 2D Keypoint Correction:
• Adopted SAM for hand segmentation, correcting keypoints on fisheye grayscale images for realistic hand depictions.
• 4. Depth Estimation & Integration:
• Integrated depth estimation with Qualcomm EVA; achieved 3ms depth for 720p images, reducing MPJP error to 6 pixels.
• 5. Gesture Recognition & Jitter Reduction:
• Applied Kalman filter, 1 Euro filter, and averaging for jitter correction in a 30 FPS pipeline, ensuring 2 pixels MPJP.
• 6. Software Testing & CI/CD:
• Tested with Gtest for module-level precision. Assessed joint and hand finder metrics.
• Enhanced workflows with Jenkins for faster CI/CD.

Apr 2022 – Jan 2023 · 9 mos · Bengaluru, Karnataka, India · On-site

• 1. Commercial of JumpShot on Samsung Devices:
• Led the design and integration of a real-time action recognition pipeline for the "Jump Shot" mode in the Samsung S22 camera app and subsequent series.
• Optimized the architecture to achieve an efficient 5ms runtime, enhancing on-device performance for immediate and responsive user experience.
• 2. Single Human Pose Estimation & Model Efficiency:
• Engineered a lightweight and efficient model using DEKR/HRNET and MobileNET for 16 key body point detection, tailored for on-device computations.
• Secured an AP0.5 score of 98% across Crowd Pose, COCO, MPII, and self-annotated datasets, ensuring high accuracy while maintaining swift processing.
• 3. Multi-human Pose Estimation & Tracking:
• Established recognition capabilities for up to 32 individuals simultaneously, integrating person tracking for continuous and coherent analysis.

Aug 2021 – Apr 2022 · 8 mos · Bengaluru, Karnataka, India · On-site

• 1. EMTC Enhancement & Optimization:
• Expanded EMTC capabilities from 2 to 4 narrow bands using C/C++, achieving:
• Doubled user base.
• Tripled range.
• Boosted uplink/downlink throughput speeds.
• Fortified security with an improved hopping algorithm.
• 2. Channel Design & Seamless Handovers:
• Engineered repetition and aggregation levels in ambiguous intervals on the MPDCCH channel, ensuring fluid connection handovers.
• 3. Leadership in SATP PoC Development:
• Led a 2-member team in developing a proof-of-concept (PoC) for the Stand Alone Test Platform (SATP) for the EMTC module using the Google Test Suite on a Linux server.
• This PoC, built using Python, Bash, C++, and MongoDB, was subsequently integrated by other teams as a foundational testing platform, streamlining debugging processes.
• 4. International Collaboration & Field Support:
• Worked alongside Samsung Vietnam and Samsung Research South Korea for:
• Field support.
• Testing.
• Log analysis related to 4G EMTC code releases, patches, and Product Lifecycle Management (PLM).

May 2019 – Jun 2019 · 1 mo · Hyderabad, Telangana, India · On-site

• 1. Comprehensive Deep Learning Coursework:
• Completed an extensive 2-month training, encompassing a wide range of topics from machine learning to artificial intelligence.
• 2. Emotion Recognition Project:
• Engineered an emotion recognition system using CNN, LSTM, and fast Fourier transform.
• Achieved 78% accuracy on the CREMA-D dataset, processing 5 hours of gender-equal monolingual data.
• 3. Helmet Detection for Traffic Control:
• Developed a helmet detection model using a modified YOLO algorithm on a CNN dataset provided by IIT Hyderabad.
• The model was successfully deployed by Hyderabad's traffic control for real-world applications.