Rohan Vayase

Hi.Start here

I build applied ML systems, then test whether they actually work.

I’m finishing an MS in Computer Science at Arizona State (May 2027). My favorite moment in any project is when a number looks too good, because that’s usually where the bug is. Chasing that feeling found a data leak in my vehicle detector and a reward bug that made my RL agent think switching lights was free. I’m looking for a full-time role building applied AI/ML systems that hold up outside the notebook.

Email meResume (PDF)LinkedIn
From my notebook: vehicle detector, YOLOv8n on UA-DETRAC traffic footage
  1. mAP50, random frame splitLeak: all 60 video clips were in both train and validation0.909
  2. mAP50, re-split by video clip0.838
  3. mAP50, clip split, 30 epochs0.907

Read the full write-up

Rohan Vayase

Not AI-generated. Validated on a held-out set of friends.

Selected work3 projects, 1 honest failure

Adaptive traffic signal control with RL and a YOLO vehicle detector2026

A simulated 4-way intersection in SUMO, controlled three ways: a fixed timer, a rule-based adaptive controller, and a DQN agent. A separate YOLOv8 detector, fine-tuned on real traffic-camera footage, is the perception front end.

Result
The rule-based controller cut average wait 66–73% across scenarios, including after adding left and right turns. The detector reached 0.907 mAP50 on video clips it never saw in training.
What didn’t work
Five DQN versions never beat the rule-based controller overall. Throughput, fairness and switching cost traded off against each other under every reward I tried.
Python, SUMO, TraCI, Gymnasium, stable-baselines3, YOLOv8, PyTorch
Read the case studyCode on GitHub
North–south greenQueued 0 NS, 0 EWThrough 0Avg wait 0.0sLive: a simplified version of the adaptive rule. Green ends when its queue clears and the other side is waiting. Click it for rush hour.

Document summarizer and RAG chatbot2025

Upload PDFs, Word files or text, build an embeddings index, then summarize the documents or chat with them. Every answer cites the chunks it was grounded in.

Result
End-to-end pipeline from ingestion to cited answers, served over FastAPI with a Streamlit UI. The LLM can be swapped between a local Hugging Face model and the OpenAI API.
What didn’t work
Retrieval quality has not been measured yet. Without a labeled question set, I can’t say how often the right chunk is retrieved.
Python, FastAPI, FAISS, Sentence-Transformers, Hugging Face Transformers, OpenAI API, Streamlit, Docker
Read the case studyCode on GitHub
DocumentsPDF, DOCX, TXTChunksclean and splitEmbeddingsSentence-TransformersFAISS indexnearest-neighbour searchTop-k chunksfor each questionLLM answercites [doc_id:chunk_id]
How a question becomes a cited answer.

Flower species recognition with CNNs2024

A convolutional neural network that classifies flower species from photos and live camera input, built as undergraduate research.

Result
Published in the International Journal for Multidisciplinary Research (IJFMR).
Python, TensorFlow, CNNs
Read the paper

Earlier projects

  • Image classification API. ResNet fine-tuned on CIFAR-100, served over FastAPI with a Streamlit front end. About 79% top-1 and 96% top-5 accuracy. Code
  • Real-time face mask detection. MobileNetV2 classifier with OpenCV face detection on a live webcam feed. Code
  • Chess bot. Minimax search with alpha-beta pruning and adjustable depth, played in the browser through p5.js. Code

Experience

Software Developer

PEOCIT Software Solutions, Pune, IndiaJun – Dec 2024
  • Built the user interface and core application logic for a stock-market portfolio management web app in C# and .NET, working directly with the client and a designer.
  • Integrated real-time stock prices from third-party APIs to power live portfolio views.
  • Reached a 97% success rate in real-time data sync by adding retry logic and error handling.

Data Analyst Intern

Poona Switchgears, Pune, IndiaMay 2023 – May 2024
  • Compiled and cleaned manufacturing data from multiple production lines with Python and pandas.
  • Analyzed it to find inefficiencies in electrical-panel assembly workflows, and presented the findings to engineering leads to inform process improvements.

Education

M.S. Computer Science

Arizona State University, Tempe, ArizonaAug 2025 – May 2027
  • Coursework: Statistical Learning Theory, Data Mining, Data Visualization, Semantic Web Mining, Knowledge Representation and Reasoning, Software Verification and Testing.

B.Tech Computer Science and Engineering, CGPA 8.36/10

MIT World Peace University, Pune, IndiaAug 2021 – May 2025
  • Published “Flower Recognition System using CNN” in IJFMR (Nov 2024). Photography Head of MIT-WPU Shutterbugs, the media club. Volunteered at Khelghar, an NGO for underprivileged children, teaching mathematics to 8th and 9th graders.

Tools

ML
PyTorch, TensorFlow, stable-baselines3, Ultralytics YOLOv8, scikit-learn, Hugging Face Transformers, Sentence-Transformers
LLM and retrieval
RAG pipelines, FAISS, OpenAI API, map-reduce summarization, tool-use agents
Languages
Python, C#, C++, Java, JavaScript, SQL
Engineering
FastAPI, Flask, .NET, REST API design, Docker, MongoDB, Git, React, Node.js
Simulation and data
SUMO, TraCI, Gymnasium, pandas, NumPy, Streamlit
Off the clock
Photography, and explaining overfitting to anyone who will stand still long enough