Engineering
Scalable
Intelligence
Building robust LLM pipelines, neural network architectures, and zero-downtime infrastructure for global-scale applications.
Technical Skillset
[Core Competencies]
System History / Log
[Experience Record]
IT Engineer - Software
[FEB 2026 — PRESENT]ORG: Keysight Technologies
Promoted from Intern - Tech 1. Built an AI-powered cloud migration assistant integrating AWS Lambda, S3, SMTP, and LLM workflows, and contributed to RAG pipelines, observability, and AI-Ops within the DigitalX Site Reliability Engineering domain.
AI ML Intern
[NOV 2025 — FEB 2026]ORG: Ubiqtech
Researched and implemented Audio LLM architectures for advanced speech processing pipelines.
AI Engineer (Part-time, Remote)
[SEP 2025 — JUN 2026]ORG: Spark And Nudge
Built an LLM-powered chatbot from scratch using the Claude API and a RAG pipeline for resume review and personalized career guidance, integrating Ollama-based workflows for contextual recommendations.
Technical Engineer Intern
[OCT 2022 — DEC 2022]ORG: Geeks for Geeks
Contributing to technical content engineering and platform workflows.
Academic Credentials
[Education History]
Indira Gandhi Delhi Technical University for Women
M.Tech, CSE (AI) · GPA 8.0
Maharaja Surajmal Institute of Technology
B.Tech, Information Technology · GPA 8.7
Research
[Thesis_Output.paper]
Predicting Non-Motor Symptom Progression in Parkinson's Disease Using a Sparse Temporal Transformer on Longitudinal PPMI Data
Ekta Dua · Anjali Lathwal · Jagrati Singh — IGDTUW
Non-motor symptoms in Parkinson's disease — cognitive, mood, sleep, autonomic, and olfactory — are notoriously hard to forecast because clinical visits are sparse and irregular. This work adapts STraTS, a sparse transformer architecture, to jointly forecast 172 item-level questionnaire variables across eight validated non-motor assessments from the Parkinson's Progression Markers Initiative (PPMI), encoding every observation as a (variable, value, time) triplet instead of relying on fixed-interval interpolation. On an independent cohort of 552 patients, the model explains 64.4% of the variance in next-visit symptom scores, with accuracy improving as more longitudinal history becomes available — a step toward personalized, longitudinal risk monitoring in PD care.
deployment:name: neural-optimizer-v2replicas: 3strategy:type: RollingUpdatemaxSurge: 1containers:- image: internal-repo/llm-orchestrator:latestresources:limits:nvidia.com/gpu: 2livenessProbe:httpGet:path: /healthzport: 8080
Featured Deployments
Archive_System_Recent
Lumi — Voice AI Agent
Built a browser voice assistant end to end: faster-whisper speech recognition, hybrid retrieval (MiniLM embeddings in ChromaDB plus keyword matching) chunked by heading with a relevance cutoff to stop the model inventing details, and Kokoro text-to-speech — running on a laptop CPU at roughly 3-4 seconds per spoken turn. Moved order logic out of the LLM into deterministic code, with a catalog validating every change and a read-back confirmation before placing an order. Dockerised for Hugging Face Spaces.
Tweet Emotion Detection
Built a recurrent neural network using TensorFlow and custom tokenizers to classify sentiment in social media data at scale.
DCGAN Digit Generation
Implemented a Deep Convolutional Generative Adversarial Network in PyTorch for synthetic handwritten digit generation.
Face Mask Detection
Built a real-time computer vision system to detect faces and identify mask usage from live video streams and static images.
Initialize Comms
Ready to integrate? Reach out to discuss system synergy, research collaborations, or engineering partnerships.