🧭 About Me

Final-year B.Tech Computer Science (Data Science) student at SVKM's NMIMS University, Hyderabad, building research-grounded AI systems that ship.

role:        AI/ML Engineer & Speech AI Researcher
focus:       Speech Recognition · Multimodal ML · Computer Vision · Explainable AI
approach:    Research depth + measurable engineering outcomes
currently:   Voice-preserving TTS · Open-vocabulary defect detection · MLOps

🛠️ Tech Stack

Languages ML / DL / GenAI

Data & Analysis Deployment & Cloud Web

🚀 Featured Projects

🎙️ Svara — Personalized Voice-Preserving Speech Correction

Speech-correction pipeline for cleft lip/palate patients using fine-tuned ASR.

• Fine-tuned Whisper-small with LoRA/PEFT, using speaker-independent, cleft-type-stratified splits
• Reduced held-out WER from 80.5% → 16.1% across three fine-tuning iterations (51/54 files perfect)
• Used dynamic per-batch SpecAugment to fix phoneme-level errors
• Extending to voice-preserving TTS (Coqui XTTS v2, OpenVoice), benchmarking on UltraSuite & TORGO

Python PyTorch Whisper HuggingFace LoRA/PEFT SpecAugment

---

🛡️ OmniSafe — Multimodal Threat Detection Platform

Late-fusion system combining visual artifact detection, semantic analysis, and explainability.

• Architected a late-fusion architecture: PyTorch GAN artifact detector + fine-tuned DistilBERT
• Added SHAP-based explanations for real-time predictions
• Optimized with ONNX Runtime + INT8 quantization: 250MB → 65MB, sub-2s latency
• Reached 0.89 Macro-F1 after resolving class imbalance

Python PyTorch DistilBERT GANs SHAP ONNX Runtime INT8 Quantization

---

🔍 Open-Vocabulary Few-Shot Defect Detection (In Progress)

Software-only inspection system for MSMEs that learns new defect classes from a handful of images.

• Frozen OpenCLIP ViT-B/16 backbone + trainable adapter + appendable prototype memory
• Learns new defect categories from just 5–10 example images
• Benchmarking CPU-only latency/memory/accuracy against WinCLIP, AnomalyCLIP, PatchCore
• Investigating ONNX INT8 quantization for edge deployment

Python OpenCLIP CLIP PyTorch ONNX Runtime Few-Shot Learning

---

📦 Saral — Distributor Management App (Client Work)

Flutter app independently built and deployed for Bajrang Distributors.

• Replaced diary-based order tracking with a digital workflow, cutting manual entry errors
• KPI dashboards for stock, customer financials, and order history
• Extending to a multi-user AWS backend

Flutter Dart Firebase REST APIs AWS SQL

📄 Publication

Sharma, A. et al. — Crime-Intel Companion: A Tamper-Proof Mobile Framework for AI-Driven Crime Scene Investigation Presented at the 2026 International Conference on Intelligent Computing and Information (ICICI), KIET

🎓 Education

Institution	Program	Duration	Score
SVKM's NMIMS University, Hyderabad	B.Tech CS (Data Science)	2024 – 2027	CGPA 3.70/4.00
Jaya Prakash Narayan College of Engineering	Diploma, Electronics & Communication	2021 – 2024	GPA 9.8/10.0
Brilliant Grammar High School	SSC	2021	GPA 10.0/10.0

🏆 Achievements

• 🌟 Selected Participant, ACM India Summer School 2026 — Systems for ML (Women Only), VIT Vellore
• 🤖 Top 50 nationally, RoboWeek 3.0 (NIT Hamirpur) — built PlantOra
• 📢 Presented research at ICICI 2026
• 🧑‍💼 Head, Institution's Innovation Council, NMIMS Hyderabad (Aug 2025 – Present)
• ⚠️ Risk Manager Lead, TechFiesta'25 · Documentation Head, Kaunyak Tech Fest

📜 Certifications

📊 GitHub Stats

📬 Let's Connect

Building reliable AI systems with research depth and engineering impact.