mayan

Resume GitHub LinkedIn

$mayan25sharma@gmail.com

VIT Chennai '27 CGPA 8.65 ex-DRDO AWS SA-A

Work

projects

FlowConX
classifies encrypted network traffic by flow shape alone.

no deep packet inspection, just packet sizes, timing, and byte rates.

90.4% accuracy at 13.65 ms per flow across 112,000 labelled flows.
source →
Tunnel
reorders Apple Music playlists for smoother listening.

on-device Core ML embeddings on the Apple Neural Engine, plus a per-playlist ranker trained in seconds.

cuts benchmark tempo jump ~55% and genre distance ~45% against a shuffled baseline.

no network calls, no cloud inference, everything runs locally.
source →
Enhanced HQ-LSTM
benchmarks quantum, hybrid, and classical forecasters.

solar PV forecasting, comparing sequence models head to head.

the code behind the IET, BICET 2025 paper.
source →
Bellwether
in progress: real time financial intelligence platform.

temporal graph risk scoring under a multi agent LLM layer.

Kafka and Neo4j underneath.

experience

DRDO
deployed a real time aircraft recognition system.

for automated ground equipment selection, improving operational safety under field constraints.

designed a two stage YOLOv8 vision pipeline for detection and classification, optimised for low-latency inference.

built a custom auto-annotation model that labelled 10,000+ images across aircraft classes, cutting dataset prep time.

passed DRDO lab evaluations as the most reliable model deployed for the use case.
VLSI Design Centre
automated digital filter design, start to finish.

a Python system that cut manual design time 50%.

numerical optimisation over 100+ parameters, improving output efficiency 25%.

an end-to-end code generation pipeline supporting 10+ design variants.

research & recognition

IET, BICET 2025 first author, Enhanced Hybrid Quantum-LSTM for Solar PV Forecasting
AWS Certified Solutions Architect, Associate
Nokia CREONIX 2025 winner

Stack

languagesPython, C++, SQL
visionYOLO, DETR, ViT, OpenCV
graphs, sequencesGNNs, Temporal Graph Networks, LSTMs, Transformers
llmsRAG, fine tuning (SFT, RLHF, GRPO), context engineering, evaluation
agentsLangGraph, LlamaIndex, smolagents, FAISS, ChromaDB
frameworksPyTorch, scikit-learn, Hugging Face (Transformers, TRL, Datasets, Accelerate)
mlopsDocker, Kubernetes, Terraform, GitHub Actions, GitLab CI, MLflow, DVC, Airflow
observabilityPrometheus, Grafana, Weights & Biases, Langfuse
cloudAWS, Solutions Architect Associate
dataPandas, NumPy, feature engineering, statistical analysis, Matplotlib
toolsGit, Linux, Jupyter

FAQ

What are you looking for?

Full time roles starting 2027: applied ML and research engineering, GenAI and LLM engineering, or ML infrastructure. Open to internships before then. Based in India, happy to relocate.

What are you building right now?

Bellwether, a real time financial intelligence platform that pairs temporal graph risk scoring with a multi agent LLM layer. Kafka and Neo4j underneath, a fine tuned small model doing the reasoning, deployed on EKS with the full observability stack. It is my final year project and the most end to end thing I have attempted.

What was the hardest thing to build?

FlowConX. Classifying traffic you cannot decrypt means every signal has to come from timing and shape: packet sizes, inter arrival times, jitter, byte rates. Getting a contrastive objective to produce clean, well separated embeddings from that, across three datasets collected under completely different conditions, took far more iteration than the architecture did.

How do you learn?

Mostly by reading source and rebuilding things. Hugging Face courses, Hands-On ML, and a lot of documentation. If I cannot reimplement it, I do not consider it learned.

How do I reach you?

Email is best, mayan25sharma@gmail.com. I am also on LinkedIn.