Specializing in Computer Vision, Uncertainty Estimation, and Production ML Pipelines (FastAPI, PyTorch,
YOLOv5, React). Formulating models from research to containerized edge deployment.
I work with real-world, imperfect data rather than clean benchmark datasets. My workflow focuses on preprocessing, class balancing, normalization, augmentation, and preparing reliable inputs before training.
3,000+ Annotated Images
Class BalancingData Augmentation
01
Engineering Background
I am a B.Tech Computer Science (AI & ML) student at Universal AI University building reliability-first
computer vision and applied AI systems.
My work focuses on bridging raw validation accuracy and production deployment. That means addressing model
uncertainty, class imbalances, edge latency, and human-in-the-loop auditability as core engineering
requirements.
Beyond engineering
Group Leader, SMART Project — led AI education sessions at rural government schools.
Member, Research Vertical & AI Cell, Universal AI University; SIH 2025 participant.
Research paper: "Harnessing AI for Sustainable Packaging" — literature review & data analysis.
Organizer & Graphic Designer — AI Odyssey Tech Event and Hawkthon Hackathon.
02
Featured Case Studies
RAMS Digital — Internship Project
Industrial Rack Damage Detection System
One of five edge-deployed computer vision systems built for warehouse
safety at RAMS Digital — flagging structural racking defects before they become safety
incidents.
Rack racking defects (arrangement issues, structural damage, corrosion/cracking, protector
damage) are missed on episodic manual inspection and carry collapse risk.
Engineering
Owned the dataset layer: hand-sorted ~11,000 raw images into a 4-class taxonomy co-designed
with RAMS's engineering team, carried through 5 dataset versions. Benchmarked YOLOv5m/l and
YOLOv8m/l detectors.
Result
Best result: 76.8% mAP@0.5 (YOLOv5m). A later reproduction on a frozen data
snapshot returned only 67.9% — a gap traced to an unpinned dependency, documented rather than
hidden, and now standing practice for every run.
A medical imaging classifier that also learns to recognize when its own
predictions can't be trusted — flagging uncertain cases for human review instead of guessing.
Status
Currently under conference review. Full architecture, results, and an interactive demo will be
published here once the paper is out — check back soon.
PyTorch Computer Vision Uncertainty-Aware AI
Full Case Study Coming Soon
RAMS Digital — Internship Project
SALT: Structural & Architecture Layout Tool
IS-code compliant multi-sheet DXF/SVG/PDF drawing packages for water
infrastructure, generated from natural language or a structured form — replacing 3–10 days of
manual drafting. Shipped first commit to production in 25 days.
Prompt-to-DXF CAD Wireframe
Engine
[Input Prompt]: "Generate a
structural layout DXF drawing compliant with relevant engineering code standards"
ezdxf 28-Layer Vector
Blueprint
Problem
Drafting water-infrastructure packages (up to 28 A1 sheets) by hand doesn't scale, and
conventions drift between engineers with no automated QA before delivery.
Engineering
A 16-step pipeline covering five project types — balancing tanks, treatment plants, elevated
reservoirs, jackwells, pump houses. IS 1172 mass-curve hydraulic sizing; drawing decisions
calibrated against 88 real reference drawings rather than hard-coded defaults; an LLM composes
sheet layout only, while deterministic Python executes the geometry.
My Contribution
Set up the SALT repo and mined 88 reference DWG/DXF drawings into layer, text-height, and hatch
conventions. Built the water-treatment-plant precedent library (38 drawings across 9 projects)
and 10 parametric unit assemblies. Implemented a 26-check automated QA module with a title-block
binder enforcing 55 canonical layers, and refactored a 4,357-line router into modular
per-type routers.
Public repository of NLP experiments — text processing pipelines, transformer
fundamentals, tokenization benchmarking, and applied exercises.
Python NLP Jupyter
03
Professional Experience
Artificial Intelligence Intern — RAMS Digital
Apr 2026 – Jul 2026
Two tracks across twelve weeks at RAMS Digital: five edge-deployed computer
vision systems for warehouse safety, and SALT, a CAD automation platform for water
infrastructure.
Owned the dataset layer for rack-defect detection: designed a 4-class taxonomy, hand-sorted
~11,000 raw images across 5 dataset versions. Best result 76.8% mAP@0.5 (YOLOv5m).
Unified 7+ public PPE datasets via an explicit class-remapping table into a 32,269-image
corpus; corrected severe class imbalance through staged rebalancing.
Converted trained detectors through ONNX export and INT8 quantization, verifying fidelity at
each stage, and deployed to a Realtek AMB82-mini NPU board (89.7% mAP@0.5).
Contributed to a Material Handling Equipment safety interlock — fail-safe brake logic and the
decision to isolate face-authentication from the on-device detector.
On SALT: mined 88 reference CAD drawings into layer/hatch conventions, built a 38-drawing
precedent library and 10 parametric unit assemblies, and implemented a 26-check automated QA
module.
PyTorch YOLOv5 / YOLOv8 ONNX / INT8 FastAPI ezdxf
Frontend Web Development Intern — CODTECH IT
Solutions
May 2025 – Jul 2025
Shipped four modular front-end web applications focused on responsive UI components
and real-time state management.
Built real-time chat application utilizing React and WebSockets with stateful history persistence.
Developed interactive quiz application with dynamic question loading and instant feedback.
Constructed multi-page e-learning interface with video embedding and user progress tracking.
React JavaScript Node.js Git
04
Technical Arsenal
Deep Learning & CV
PyTorchYOLOv5 / YOLOv8ResNet-50OpenCVONNX / INT8 QuantizationEdge AI DeploymentExplainable AIUncertainty Scoring
