Aman Dalmia

amandalmia18@gmail.com | +917003060871
http://amandalmia.weebly.com | http://linkedin.com/in/aman-dalmia | https://github.com/dalmia

EDUCATION
IIT GUWAHATI | B.TECH, ELECTRONICS & COMMUNICATION ENGINEERING Guwahati, India | May 2018
GPA: 3.24/4
BHARTIYA VIDYA BHAVAN | CBSE BOARD, SCIENCE Kolkata, India | May 2014
GPA: 93.4% in 12th | 10/10 in 10th

EXPERIENCE
WADHWANI INSTITUTE FOR AI | RESEARCH FELLOW Mumbai | July 2018 - Present
• Developed an early pest-detec on framework using field images collected and annotated by our team, using a
deep learning based object-detec on framework comparing different single-stage and two-stage detectors.
• Proposed a new object algorithm adds scale selec on to the learning process itself. Achieved a mean F1 score
of 91.2% across several classes of worms.
• Building an end-to-end pipeline to iden fy low-birth weight babies using smartphone-based images and
videos, using a learned deformable model capturing the latent subspace of realis c baby shapes.
• Conducted the internship hiring process for the next year and successfully hired 6 interns from various IITs
among a pool of 250 applicants.
HYPERVERGE | DEEP LEARNING INTERN Bangalore | May 2018 - Present
• Worked on the problem of change detec on where the inputs are the images of a place at two different
mestamps and the expected output is the change that happened in the me dura on.
• Designed the seman c segmenta on pipeline and used various architectures like U-Net, V-Net and
DeepLabv3+ combined with trying out different loss func ons, ac va on func ons like PReLU and Maxout,
variants of the original architectures, op mizers like SGD & Adam along with various regulariza on techniques
like Data Augmenta on (geometrical/color) to monitor and improve the segmenta on map to obtain a
pixel-wise accuracy of more than 90% using data provided by the client.
• Modularized the en re codebase by following object-oriented programming principles and used
mul processing along with queues to ensure a fast and efficient implementa on of various steps like
inference, finding op mal threshold, etc.
• Used the Keras framework for the project with TensorFlow backend. Other major libraries used were
OpenCV, Numpy, Matplotlib and scikit-learn.
NIRAMAI | RESEARCH INTERN - DEEP LEARNING | PUBLISHED IN QIRT Bangalore | Dec 2017
• Developed the pipeline for Seman c Segmenta on on thermographic images to get a final dice-coefficient of
0.83, implemen ng various CNN architectures like U-Net, V-Net, Convolu onal Sliding Window and
mul -channel cascaded CNN, using Keras with TensorFlow backend.
• Wrote custom layers, experimented with different metrics and u lized Bayesian Op miza on for
Hyperparameter tuning. Our paper was published at the Quan ta ve Infrared Thermography (QIRT)
Conference, 2018 held in Berlin.
QUALCOMM | SOFTWARE DEVELOPMENT INTERN | PROJECT LINK Bangalore | May 2017 – July 2017
• Developed a real- me car parking system model applying a Convolu onal Neural Network trained on the
PKLot dataset with an accuracy of almost 94% using Transfer Learning.
• Proposed a novel algorithm for internal naviga on using Li-Fi and preven ng car-the using Machine Learning
& Image Processing.
SCHOLAR MATE | CO-FOUNDER | APP LINK Guwahati | Jan 2016 – May 2016
• Led a team of 5 people and visited 50+ schools taking user feedback to finally launch in 1 school, with the aim
of bringing together teachers, parents and students on a single pla orm (100+ downloads in the first week).
• Contributed in developing the Android Applica on (Java, XML) and the back-end (PHP, MySQL).
TWANGO | SOFTWARE DEVELOPMENT INTERN - ANDROID | PROJECT LINK Delhi | Dec 2016
• Developed the first version of the Android Applica on (currently 50k+ downloads) (5k+ LOC), a reliable
pla orm for like-minded singles to connect.
• Collaborated with mul ple teams (design, marke ng, psychology) to determine implementa on methods and
project melines, refactoring the en re code base to match the needs as per the user feedback.

PROJECTS
SPLICE SITE PREDICTION USING DEEP LEARNING Prof. Ashish Anand, Dept. of CSE, IIT Guwahati
Employing Machine Learning for understanding the genome, modelling each DNA sequence as a sentence, our
implementa on of Long Short-Term Memory (LSTM) surpasses the state-of-the-art results with an accuracy of
98.3% on unseen sequences for splice-site predic on. We further perform a quan ta ve as well as qualita ve
analysis of our model and infer that the model focuses on parts that are important for splicing as men oned in the
literature. This helps us to iden fy mo fs that are cri cal for splicing. Also, our present results are restricted to
fixed-length sequences and so, we’ve started exploring Autoencoders to extend our analysis to larger sequences.
Paper submi ed to the Bioinforma cs Journal.
ADVANCE LANE LINES Link to Github
Developed a sophis cated pipeline to detect lane lines in videos robust to shadows and curving roads, completed
as part of the Self-driving Car Nanodegree Program by Udacity. The pipeline consists of several steps: camera
calibra on to obtain distor on coefficients for undoing the distor on in each frame, perspec ve transform to
obtain a bird-eye’s view, applying color and gradient-based thresholds and using a sliding window based approach
for fi ng quadra c polynomials to pixels corresponding to the lane lines.
FINDING LANE LINES ON THE ROAD Link to Github
Built a computer vision pipeline to detect lane lines in videos using Python and OpenCV. This was completed as
part of the Self-driving Car Nanodegree Program by Udacity. The pipeline consists of conver ng images to
grayscale, adding Gaussian blur, then performing Canny Edge detec on and finally, apply Hough transform to get
the detected lane lines. There is an extrapola on step needed to return just one lane line for either sides.
AN ACTOR-CRITIC METHOD FOR CONTROL IN CONTINUOUS ACTION SPACE Link to Github
In this environment, there are 20 agents and each agent is a double-jointed arm that can move to target loca ons.
A reward of +0.1 is provided for each step that the agent’s hand is in the goal loca on. Thus, the goal of the agent
is to maintain its posi on at the target loca on for as many me steps as possible. I used a modifica on of the
Deep Determinis c Policy Gradient (DDPG) algorithm and was able to solve the environment in 173 episodes
achieving an average reward score of more than +30. The linked repository contains the code and checkpoints
along with comparison with a random agent.
DEEP Q-NETWORKS FOR OPTIMAL NAVIGATION Link to Github
The project demonstrates the ability of value-based methods, specifically, Deep Q-learning and its variants, to learn
a suitable policy in a model-free Reinforcement Learning se ng using a Unity environment, which consists of a
con nuous state space of 37 dimensions, with the goal to navigate around and collect yellow bananas (reward: +1)
while avoiding blue bananas (reward: -1) with 4 ac ons to choose from: move le , move right, move forward and
move backward. It also contains implementa ons for Double DQN and Dueling Double DQN along with their
performance comparison.
DEEP SEMANTIC KERNEL (DESK) Prof. Amit Sethi, Dept. of EEE, IIT Bombay
Applying transfer learning to accurately classify images from the Omniglot dataset, which has only 20 examples per
class, using a Siamese Neural Network in associa on with SVM.
QUORA QUESTION PAIRS ON KAGGLE https://dalmia.github.io/Quora-Question-Pairs/
Trained the Gated Recurrent Unit (GRU) variant of RNN in combina on with Random Forest for predic ng
duplicate ques on pairs to land in the top 25% on Kaggle among 3000+ teams with a cross-entropy loss of 0.21.
CANCER RECURRENCE PREDICTION
Performed a literature survey on various Convolu onal Neural Network (CNN) architectures used for
classifying images & developed a transfer learning model for predic ng Cancer Recurrence using
whole-slide ssue images of pa ents.

PUBLICATIONS
[1] Aparajita Du a, Aman Dalmia, Athul R, Kusum Kumari Singh, Ashish Anand: Inference
of splicing mo fs through visualiza on of recurrent networks; Submi ed to Bioinforma cs
[2] Aman Dalmia, Siva Teja Kakile and Geetha Manjunath: Exploring Deep Learning
Networks for Tumour Segmenta on in Infrared Images; Published in QIRT 2018, Berlin
[3] Ameya Godbole, Aman Dalmia and Sunil Kumar Sahu: Siamese Neural Networks with Random
Forest for detec ng duplicate ques on pairs; arXiV: h ps://arxiv.org/abs/1801.07288

SKILLS
PROGRAMMING LANGUAGES Python, C++, Java, C, SQL, HTML, PHP
ARTIFICIAL INTELLIGENCE Machine Learning, Deep learning, Computer Vision, NLP
TOOLS / FRAMEWORKS PyTorch, Keras, TensorFlow, Numpy, scikit-learn, Matlab, Git, Github

OPEN SOURCE
SCIKIT-LEARN | ML LIBRARY IN PYTHON | WEBSITE | GITHUB Oct 2016 - Feb 2017
• Resolved issues involving documenta on, API changes, bug fixes, chunking & memory enhancements.
• Reviewed pull requests by other contributors and gave them feedback to improve their code quality.
• All my contribu ons can be seen here.

DEEP LEARNING BOOK SUMMARIES | AUTHOR | LINK TO REPOSITORY

VOLUNTEER EXPERIENCE
Co-founder and Mentor IITG.ai Present
City Ambassador AI Saturdays Feb 2018
Open Source Developer scikit-learn Nov 2016 - March 2017
Workshop Facilitator Google Developers’ Group, IITG Feb 2017

AWARDS
UDACITY SELF-DRIVING CAR NANODEGREE MARCH 2019
Selected as a Udacity KPIT Tech Scholar for Udacity’s Self-driving Nanodegree Program.
ICNDE MARCH 2018
Selected to a end the Interna onal Conference on Networked Digital Earth conducted by IIT Kharagpur.
1ST PLACE AT HACK IN THE EAST JAN 2018
Created a prototype to demonstrate the idea of ”Waste disposal using Deep Learning on Drone Images”
9TH PLACE IN GS QUANTIFY ’18 AMONG 270+ TEAMS ALL OVER INDIA OCT 2017
The challenge involved 3 ques ons from Computer Science, Machine Learning & Quant. I was responsible primarily
for the ML ques on and contributed partly to the Quant one as well,
INTERNATIONAL FELLOWSHIP BY FAST.AI OCT 2017
MIT MEDIA LABS WORKSHOP JAN 2017
Among the 9 candidates who were selected from all over India to a end the workshop in Mumbai.
ANDROID DEVELOPER NANODEGREE SCHOLARSHIP MAY 2016
Received the scholarship offered by Google & Tata Trusts for Udacity’s Android Developer Nanodegree Program.
2ND PLACE, STARTUP ASSAM FEB 2016
Runners-up of the B-plan compe on conducted by Startup Assam, among 34 other teams.
1ST PLACE AT CODE IO, IIT GUWAHATI JAN 2016
Developed an android app (Even fy) along with a browser extension that directly sends specific details (chosen by
the user) of the ins tute mail to the app using Firebase Cloud Messaging.

INVITED TALKS
BAYESIAN DECISION MAKING MUMBAI, SEPTEMBER 2018
Invited as one of the speakers for a code-along session a ended by more than 100 people which was organized by
DataGiri, a community of more than 15,000 data science enthusiasts. Slides for the session can be found here.
IIT GUWAHATI ALUMNI MEETUP MUMBAI, NOVEMBER 2018
Gave a talk on ”Demys fying entrepreneurship, doing something you love and giving back” at the IIT Guwaha
alumni meet up in Mumbai.

COURSEWORK
• Algorithms & Data Structures • Pa ern Recogni on & Machine Learning (10/10)
• Informa on Theory & Coding (10/10) • Deep Learning for NLP • CNN for Visual Recogni on
• CNN for Visual Recogni on • Image Processing • Deep Learning Specializa on
• Opera ng Systems • Communica on Networks • Computer Vision*
• Reinforcement Learning (UCL)* • Theories of Deep Learning (Stanford)*
• Economics of Climate Change* • Solar Energy Conversion Technology*