"Raise yourself by yourself; do not let yourself fall."
Machine learning engineer focused on LLM alignment, concept drift detection, and on-device inference. B.Tech CSE · SRMIST Chennai · 2023–2027.
shoryamishra65@gmail.com · linkedin · X/twitter
Large language model alignment and evaluation · Concept drift detection in clinical NLP pipelines · Efficient neural architectures and learned sparsity · On-device sensor-fusion ML systems · Retrieval-augmented generation
Python · PyTorch · HuggingFace · BioBERT · FAISS · FastAPI
Peer-reviewed research on temporal alignment drift in production LLMs (GPT-4, Llama-2, Mistral) exposed to evolving clinical guidelines. The core contribution is a RAG pipeline with section-aware semantic chunking, BioBERT embeddings, and FAISS vector indexing designed to intercept guideline drift before it propagates to downstream inference.
- Evaluation on ConflictMedQA-Extended: 97.8% sensitivity, 88.5% specificity
- 18% reduction in false-positive drift flags via max-ensemble (BioBERT + Qwen CoT)
- Automated MLOps telemetry: KS-tests and Population Stability Index for live distribution monitoring via W&B
Python · PyTorch · Sentence-BERT · Transformers · Streamlit
An LLM output auditor targeting reward-hacking artifacts that propagate silently through SFT data pipelines. Motivated by OpenAI's documented alignment failures in RLHF training. Three detection signals are fused into a composite TicScore (0–100): n-gram lexical drift scanner, TicAutoencoder reconstruction error over Sentence-BERT embeddings (all-MiniLM-L6-v2, 384-dim), and a fine-tuned TicClassifier. Outputs structured JSON reports. Relevant to safety-critical deployment contexts: medical AI, autonomous decision systems.
Python · PyTorch · CIFAR-10 · CUDA
Novel PrunableLinear layer with learnable sigmoid gates and L1 sparsity regularisation. Validated through systematic ablation across sparsity coefficients, gate initialisation schemes, and fine-tuning schedules.
- 91.3% weight sparsity (3.47M of 3.8M parameters pruned)
- 2.6% accuracy degradation on CIFAR-10 at 11.5× compression
- Direct applicability to on-device LLM inference and edge deployment
Python · PyTorch · YOLOv8 · CUDA · OpenCV · FastAPI
Custom YOLOv8 model trained on 2,500+ annotated synthetic images (COCO-format, 25 GB) for real-time object recognition in constrained operational environments. Augmentation strategy: mosaic composition, mixup regularisation, HSV-jitter. Modular FastAPI inference backend with confidence-threshold controls.
- mAP@0.5: 0.936 · Precision: 0.967 · Recall: 0.895
- 28 FPS inference on GPU; fault-tolerant training with CUDA OOM recovery
TensorFlow · Keras · ResNet-50 · Transfer Learning · OpenCV
Transfer learning–based classifier on histopathological CT scans. Fine-tuned ResNet-50 backbone; addressed class imbalance via augmentation pipeline (rotation, elastic distortion, colour jitter). 90% classification accuracy; 35% reduction in training time versus scratch training.
AI / ML Engineer Intern — Infosys Springboard (Sep 2024 – Mar 2025, Remote)
Trained and deployed PyTorch models with Optuna/GridSearchCV hyperparameter search and k-fold cross-validation; F1 improved 12% on class-imbalanced real-world datasets. Built modular, reproducible data pipelines (ingestion → feature engineering → model training → evaluation), reducing manual preprocessing effort by 40%.
Software Engineer Intern, iOS & Systems — Infosys × Apple iOS Dev Centre, SRMIST (Apr 2025 – May 2026)
Selected in the top 4% (100 of 3,000+ applicants). Apple Swift Student Challenge 2026 Global Winner (~350 selected from 50,000+ worldwide): built AXIS, a real-time sensor-fusion application leveraging CoreMotion and HeadphoneMotionManager for on-device posture analytics — fully offline, zero network dependency. Also architected Rentiwise (live on App Store): async/await networking, in-memory caching, REST API pipeline; data-fetch latency reduced 35% at sustained 60 fps.
iOS Developer Intern — Infosys Fleet Management Systems, Mysore (Feb 2026 – Mar 2026)
Designed and shipped a Fleet Management System with real-time vehicle tracking, spatial indexing, and graph-based routing. Ranked #1 across all intern submissions company-wide. Delivered 3 production-ready features within a 48-hour sprint; cleared full QA with zero post-release defects.
| Domain | Tools & Frameworks |
|---|---|
| ML / DL | PyTorch, TensorFlow, Keras, scikit-learn, XGBoost, CUDA, SHAP, Optuna, W&B |
| NLP / LLM | HuggingFace Transformers, LangChain, OpenAI API, LoRA/QLoRA, RLHF, spaCy |
| RAG & Retrieval | FAISS, ChromaDB, BERTScore, semantic chunking, cosine-similarity retrieval |
| CV | YOLOv8, OpenCV, ResNet-50, transfer learning |
| Infrastructure | FastAPI, Docker, AWS (EC2, S3, SageMaker), Firebase, Git, CI/CD, Linux |
| Languages | Python (primary), C/C++, Java, SQL, Swift, Go, Bash |
- Apple Swift Student Challenge 2026 — Global Winner (~350 of 50,000+ worldwide submissions)
- Guidewire Hackathon — Top 6 of 5,000+ teams (GigBuddy: ML-powered parametric insurance risk engine)
- Google CodeClash 2.0 — Top 13 of 1,300+ teams
- Smart India Hackathon 2024 — National Participant (Government of India flagship programme)
- DriftShield — Peer-reviewed publication in NLP / clinical AI safety
AWS Certified AI Practitioner · AWS Certified ML Engineer – Associate · NVIDIA: Fundamentals of Deep Learning · Oracle: Generative AI Professional · SAP Certified GenAI Developer · Salesforce Agentforce Specialist
Publish technical breakdowns of NLP and ML research papers on X/Twitter — covering LLMs, mechanistic interpretability, computer vision, and reinforcement learning. 200+ posts; audience is practitioners and researchers.
Open to research collaborations, internships, and full-time roles in ML engineering and applied AI.