Download Resume
Available for opportunities

Sahil Duwal

Computer Engineer AI / ML Engineer Data Analyst

Computer Engineering student building intelligent systems at the intersection of AI, computer vision, and data science. Focused on delivering real-world impact through rigorous engineering and research.

View Projects Download CV
Sahil Duwal
▲ AI / ML
◆ Computer Vision
8+ AI Projects
10+ Certifications
3+ Years Learning
Scroll
About

Engineering the Future
One Model at a Time

I'm a Computer Engineering student, deeply invested in artificial intelligence, machine vision, and data-driven systems. My work spans from deepfake detection architectures to AI-powered travel assistants.

I believe that meaningful engineering comes from understanding both the underlying mathematics and the real-world context. Every project I take on is a step toward building systems that genuinely help people.

See My Work
◆

Focus Areas

Deep Learning, Computer Vision, NLP, Data Science

■

Currently

DataCamp Fellow · Code for Nepal · BCom Eng student

○

Location

Bhaktapur, Nepal

▲

Open To

Internships, Research Collaborations, Full-time Roles

Expertise

Technical Skills

Organized by project lifecycle stage: Data → Modeling → Deployment → Tools.

01

Data Engineering

Data collection, cleaning, annotation, and pipeline construction.

Python SQL Pandas NumPy Data Cleaning ETL Pipelines
02

Modeling & Research

Model design, training, validation, and optimization for production.

PyTorch Scikit-learn Computer Vision Deep Learning NLP Experiment Tracking
03

Deployment & Infrastructure

Production-ready model serving, containerization, and system integration.

FastAPI Docker Linux (Arch/CachyOS) Next.js Streamlit CI/CD
04

Tools & Languages

Core programming languages and essential development tools.

Git / GitHub C / C++ LaTeX HTML / CSS / JS JSON Automation Jupyter
Experience

Professional Journey

Oct 2025 – Dec 2025 Internship

Data Annotation Intern

Next Wave AI

  • Annotated images and comic covers to support computer vision model development.
  • Created structured JSON pipelines for AMS360 workflow automation.
  • Evaluated LLM security behavior through systematic prompt testing and red-teaming analysis.
  • Researched and benchmarked AI video generation tools for capability assessment.
2025 – Present Fellowship

DataCamp Fellow

Code for Nepal

  • Selected for a competitive data science fellowship program.
  • Advancing expertise in Python, data analysis, and machine learning applications.
2024 – 2025 Volunteer

Tech Support Member

Hult Prize — Khwopa College of Engineering

  • Delivered technical support for event management, presentation setups, and logistics.
  • Ensured seamless operations across multiple Hult Prize on-campus events.
Laboratory

Research & Development

Rigorous AI/ML engineering projects with documented architectures, datasets, and evaluation metrics.

AI / ML · Flagship
↗

Deepfake Detection & Localization

Developed a hybrid ViGNN–U-Net model for detecting and localizing deepfake content with high accuracy and robustness across multiple manipulation types.

Model Architecture
Hybrid ViGNN–U-Net
Dataset
Multi-source face manipulation dataset
Key Metric
85% Detection Accuracy
💡 Key Insight

Hybrid architecture combining ViGNN's frequency analysis with U-Net's spatial localization achieved superior performance over single-branch models.

PyTorch U-Net ViGNN Computer Vision Deep Learning
AI / ML · Medical
↗

TumorNet — Brain Tumor Segmentation

U-Net model for brain tumor segmentation from MRI scans, achieving strong Dice coefficient and IoU metrics with custom preprocessing pipelines.

Model Architecture
U-Net Encoder-Decoder
Dataset
BraTS MRI scans (3D volumes)
Key Metric
0.89 Dice Coefficient
💡 Key Insight

Custom augmentation pipeline including elastic deformations improved robustness on limited medical imaging data.

PyTorch U-Net Albumentations Medical Imaging
AI / ML · CV
↗

Facial Emotion Detection

Deep CNN model classifying seven human emotions from static images and real-time webcam feeds using PyTorch and transfer learning.

Model Architecture
Custom CNN (7-class)
Dataset
FER2013 + custom annotations
Key Metric
78% Classification Accuracy
💡 Key Insight

Transfer learning with pre-trained ResNet backbone significantly reduced training time and improved convergence.

PyTorch CNN Computer Vision
AI / ML · NLP
↗

PDF Q&A — Local RAG Bot

Local multi-document RAG app for semantic PDF retrieval and grounded answers with Llama 2.

Model Architecture
LangChain + Llama 2 + Embeddings
Dataset
User-supplied PDFs (Doc Vector Store)
Key Metric
<500ms Response Time
💡 Key Insight

Semantic chunking + RAG pattern eliminated hallucinations while maintaining sub-500ms latency for local deployment.

LangChain Llama 2 (Ollama) sentence-transformers Streamlit
AI / ML · Finance
↗

Fraud Detection — Transaction Security

FNN-based fraud engine with weighted BCE loss and threshold tuning for highly imbalanced datasets.

Model Architecture
Feed-Forward Neural Network
Dataset
Credit card transactions (99.8% imbalanced)
Key Metric
0.94 F1-Score
💡 Key Insight

Weighted loss with class-specific thresholding achieved 94% F1 despite severe class imbalance via SMOTE.

PyTorch FastAPI Pandas Imbalanced Learning
Productivity · Web
↗

Modern To-Do — Productivity Suite

Task manager with local LLM chat, canvas progress tracking, custom calendar, and persistent themes.

Tech Stack
Vanilla JS + Express
Features
LLM Chat, Canvas, Calendar
Key Metric
~2sec Load Time
💡 Key Insight

Local-first architecture with IndexedDB provides offline capability and sub-second task operations.

Vanilla JS Node.js Express Ollama (Phi-3)
AI / ML · CV
↗

MaskGuard — Real-Time Detection

Real-time mask detection using fine-tuned ResNet18 with MediaPipe localization and OpenCV inference.

Model Architecture
ResNet18 + MediaPipe
Dataset
Custom masked/unmasked face dataset
Key Metric
28ms Inference Speed
💡 Key Insight

ResNet18's lightweight design enabled real-time processing at 28ms/frame while maintaining 96% accuracy.

PyTorch MediaPipe OpenCV ResNet18
AI / ML · CV & NLP
↗

VisionScript — Image Captioning

End-to-end image captioning with ResNet50 encoding and LSTM decoding plus beam search generation.

Model Architecture
ResNet50 Encoder → LSTM Decoder
Dataset
Flickr8k (8K images + 40K captions)
Key Metric
BLEU-4: 0.32
💡 Key Insight

Beam search with length penalty prevented caption truncation; attention mechanisms localized salient image regions.

PyTorch ResNet50 LSTM (RNN) Flickr8k
Research

Technical Deep Dives

In-depth explorations of engineering challenges and solutions.

Optimizing Wav2Vec2 for Low-Resource Languages

2025 Intermediate

Fine-tuning speech recognition on limited Nepali datasets while maintaining acoustic robustness across dialects.

Challenge: Low-resource languages lack labeled audio. Direct fine-tuning overfits and fails across dialects.

Solution: Three-stage approach—self-supervised pretraining on unlabeled audio, supervised fine-tuning with SpecAugment, and accent-aware reweighting using SMOTE.

$L = \alpha L_{ctc} + (1-\alpha) L_{accent}$ where $\alpha=0.5$

Result: 18% WER (vs 32% baseline), 76% generalization across 5 dialects.

Wav2Vec2 Speech Augmentation

Handling Class Imbalance in Medical Image Segmentation

2025 Intermediate

Training U-Net on imbalanced datasets where tumor pixels comprise <1% of total pixels.

Challenge: Medical images have severe class imbalance; standard BCE loss biases toward background.

Solution: Focal Loss ($\gamma=2$) + Weighted Dice + patch-based sampling + post-hoc thresholding.

$FL(p) = -\alpha_t(1-p_t)^{\gamma} \log(p_t)$

Results: Standard BCE: 0.71 Dice → Focal Loss + Weighted Dice: 0.91 Dice.

U-Net Medical Focal Loss

Building Production-Grade RAG Systems

2025 Advanced

Deploying retrieval-augmented generation for accurate, grounded QA in local environments without hallucination.

Challenge: Basic RAG hallucinates when retrieval fails. Production systems need reliability metrics.

Solution: Semantic chunking + Hybrid retrieval (BM25 + dense embeddings) + Reranking layer + Confidence scoring.

$Q = 0.4 \cdot MRR(BM25) + 0.6 \cdot MRR(Dense)$

Results: <5% hallucination rate, 320ms latency (retrieval + generation). Deployed on-prem with Llama 2.

RAG LLM Production
Education

Academic Background

2021 – Present
Bachelor of Computer Engineering
Khwopa College of Engineering, IOE — Libali, Bhaktapur

Comprehensive program with focus on AI, machine learning, and software development. Engaged in hands-on labs and research-driven coursework.

Current
2019 – 2021
Higher Secondary Education (+2), Science
Khwopa Secondary School — Dekocha, Bhaktapur

Strengthened foundations in mathematics and sciences.

Completed
Recognition

Achievements

2025

Locus Pattern Verse

Advanced to Phase-2 in the national tech competition with an AI-powered rug design generation system.

2025

CodeYatra Hackathon

Participated in a national-level innovation hackathon, focusing on AI-driven problem-solving solutions.

2022–23

Hult Prize Nepal

Reached Semi-Finals with a campus team startup idea presenting an innovative social impact solution.

Credentials

Certifications

DataCamp
AI Fundamentals
2025
DataCamp
Data Skills for Business
2025
DataCamp
EU AI Act Fundamentals
2025
DataCamp
Understanding Data Topics
2025
Cisco Networking Academy
Python Essentials 1
2024
Cisco Networking Academy
Introduction to Cybersecurity
2024
Cisco Networking Academy
Introduction to IoT
2024
Programiz Pro
Learn HTML
2024
Programiz Pro
SQL Basics
2024
Programiz Pro
Learn Python Basics
2024
Contact

Let's Connect

Open to internships, research collaborations, and full-time opportunities. Reach out and let's talk.

Email
sahilduwal@gmail.com
↗
LinkedIn
Sahil Duwal
↗
GitHub
Sahahahil
↗