~/ai-engineer

Devisetty Lakshmi Srinivas

Building machine learning models that hold up beyond the benchmark.

M.Tech candidate in Artificial Intelligence and Machine Learning. I build and evaluate deep learning models in Python and PyTorch across computer vision, graph neural networks and natural language processing.

View projects Resume

Status

Open to work

Entry-level AI Engineer and Data Scientist roles

Education

M.Tech, AI and ML

Amrita Vishwa Vidyapeetham · December 2026

Latest blog

Notes on models, machines and games

On the blog

Python
PyTorch
PyTorch Geometric
Hugging Face
scikit-learn
OpenCV
Captum
pandas
NumPy
SQL
Linux
Git

01 — About

Broken runs forge better models.

I am completing an M.Tech in Computer Science and Engineering (AI and ML) at Amrita Vishwa Vidyapeetham and will graduate in December 2026. My dissertation builds a benchmark to test whether explanations from graph neural networks point to real disease signals or to scanner artefacts.

I enjoy problems that need both first-principles thinking and careful engineering. I break a vague question into testable pieces, build the simplest thing that could work, then try hard to prove it wrong. I read papers to borrow ideas, write clean and reproducible code, and treat evaluation and failure analysis as part of the work, not an afterthought.

Education

2024 – 2026 M.Tech, Computer Science and Engineering (AI and ML)
Amrita Vishwa Vidyapeetham · Expected December 2026
2020 – 2024 B.Tech, Computer Science and Engineering (AI)
Parul Institute of Engineering and Technology · CGPA 7.73 / 10

How I work

01Frame
Pin down the real question, the constraints and what success looks like.
02Research
Read papers and docs, study prior work and pick the right approach.
03Prototype
Build the simplest working baseline fast, then improve it step by step.
04Validate
Test hard: clean splits, ablations, failure cases and honest metrics.
05Ship & share
Clean code, version control, documentation and a clear write-up.

02 — Projects

Selected work.

Pixels → graphs → tokens: projects across computer vision, graph neural networks and LLMs, building toward data science and AI engineering.

M.Tech Dissertation · 2026 In progress ↗

Connectome Explanation Benchmark (CXB)

A benchmark that plants a known connectivity effect into real ABIDE I fMRI connectomes (848 subjects, 19 hospitals), so model explanations can be scored against an exact ground truth.

Showed that a train-only scanner shortcut raises the scanner share of the top explanation edges by up to 0.43, while test accuracy falls by only 0.03 to 0.07.
Proposed counterfactual site training (CF-Site): +0.15 AUPRC in disease recovery over plain training, matching the best ComBat variant without harmonizing new-site data.

PyTorch Geometric
Captum
nilearn
scikit-learn

Kaggle · 2026 In progress ↗

LLM Response Preference Prediction

Predicting which of two LLM answers a human preferred (A, B or tie) on 57K Chatbot Arena battles across 64 LLMs, evaluated on log loss.

Analyzed per-model win, loss and tie rates to understand how preferences vary across models.
Found and fixed a pipeline bug that broke Kaggle scoring, then shipped a validated baseline.
Now adding TF-IDF text features to move beyond model-identity signals.

pandas
scikit-learn
Matplotlib
Seaborn

M.Tech Thesis Completed ↗

SMOTE-Enhanced Multi-View Deep CNN for Mammogram Classification

A 2.24M-parameter multi-view (CC and MLO) CNN for benign and malignant classification on CBIS-DDSM (1,566 patients), with stratified patient-level 70/15/15 splits. Read the thesis.

Engineered a SMOTE-based view-completion method that synthesizes missing views by donor-based convex interpolation, growing complete four-view cases from 105 to 1,223.
Ran a three-tier experiment (baseline, view completion, hybrid class balancing). The final tier reached 84.02% accuracy, 92.86% recall, 85.25% F1 and 93.24% AUC.
Built the full pipeline: CLAHE enhancement, class-imbalance handling, BCE-with-logits training, learning-rate scheduling and early stopping.

PyTorch
CLAHE
SMOTE
scikit-learn

Intel Unnati Industrial Training · 2023 Completed ↗

Road Object Detection using YOLOv3

A YOLOv3 (Darknet-53) object detector implemented in PyTorch for Indian road scenes.

Implemented YOLOv3 in PyTorch, including the config parser, pre-trained weight loader and detection layers.
Wrote the post-processing pipeline: bounding box decoding, IoU, confidence thresholding and non-maximum suppression.
Ran inference on 416×416 Indian Driving Dataset images and timed it against an OpenCV DNN pipeline.

PyTorch
OpenCV
NumPy
Matplotlib

03 — Skills

Technical toolkit.

{ }Programming languages

Python
C
C++
Java
Bash
SQL

∇ML and DL frameworks

PyTorch
PyTorch Geometric
Hugging Face Transformers
Captum
scikit-learn
OpenCV

◇ML and DL techniques

Model evaluation
Fine-tuning
Transfer learning
CNNs
GNNs
Computer vision
NLP
Explainable AI
Object detection

∑Data science and analytics

pandas
NumPy
SciPy
statsmodels
Matplotlib
Seaborn
Data preprocessing
EDA
Feature engineering
Statistical analysis
Hypothesis testing
Data visualization
Survival analysis

⌘Databases and tools

MySQL
MongoDB
Git
GitHub
Jupyter Notebook
Google Colab
Linux

⟨⟩Core computer science

Data structures and algorithms
Object-oriented programming
Operating systems
Algorithm design

04 — Achievements

Certifications and recognition.

Recognition

TCS CodeVita Season 11, global rank 4,774Competitive programming

Certifications

Intermediate Deep Learning with PyTorchDataCamp
Power Platform FundamentalsMicrosoft
Crash Course on PythonGoogle

Notes and reports

05 — Coding activity

Live across platforms.

Pulled from each platform's public profile.

GitHub

@ZEROBYTE57

Live

Contributions, last year: –
Public repositories: –
Followers: –

LeetCode

– problems solved

Easy
–
Medium
–
Hard
–

GeeksforGeeks

– problems solved

Coding score: –

HackerRank

– skill badges

06 — Beyond research

Technical interests.

I write about these on my blog. For life off the grid, visit 404 Realm →

Linux env

Fedora is my flavour of Linux. I test different kernels and CPU schedulers on it, and currently run a custom CachyOS kernel alongside the stock Fedora kernel. Before settling on Fedora I also ran CachyOS, Ubuntu and Garuda.

voidsorrow@fedora:~

$ whoami

voidsorrow · ML engineer · gamer

$ tree -L 2 ~/zerobyte57

~/zerobyte57 ├── portfolio/ │ └── projects · skills ├── blog/ │ └── articles · graph └── 404-realm/ └── games · rig · discord

3 directories

Algorithms and problem solving

I write code the way algorithms are designed: define the problem precisely, reason about complexity before the first line, and choose the data structure that makes the solution almost obvious.

Explainable AI

Checking what a model actually learned: attribution methods, shortcut detection and explanations that can be verified, not just visualised.

LLMs and generative AI

How large language models are fine-tuned, evaluated and compared, from human-preference data to how a model's behaviour changes with its training.

Systems and performance

Getting more out of the hardware: kernel and scheduler tuning on Linux, efficient GPU training and finding where slow code really spends its time.

Game development

How games are built: game loops, mechanics, engines and game AI, where machine learning meets interactive systems.

Programming

Writing clean, tested code in Python, C, C++ and Java, from small tools and scripts to full implementations.

Web development

Fast, accessible interfaces in plain HTML, CSS and JavaScript. This site is hand-built without frameworks, with live data feeds and its own graph engine.

07 — Off the grid