Srijan Dawadi

The Problem

Nepal had no reliable digital platform where local service providers — electricians, plumbers, tutors, designers — could list themselves and get found by buyers. Most work was through word-of-mouth, with no trust signals, no messaging, and no secure payment layer.

My Role

Co-built the full platform as part of a 2-person IT team for PerfectLink, a registered company in Nepal. I owned the backend architecture, the AI search pipeline, and the WebSocket messaging system.

What I Built

A full-stack marketplace with a Django REST backend organized into modular apps for accounts, listings, messaging, and payments. The frontend runs Next.js with TypeScript and server-side rendering for SEO. A WebSocket server handles real-time buyer-seller messaging and price negotiation, while an AI search pipeline uses embedding generation and semantic search to surface relevant listings — so a search for "fix my pipe" finds plumbers, not zero results. The platform also includes seller dashboards, a review and rating system, and payment integration points.

Key Technical Decision

Chose semantic search over keyword search from the start — a keyword search for "fix my pipe" would return zero results if sellers listed themselves as "plumber." Embedding-based search solved this and made the product actually usable for non-technical users.

Outcome

Live at perfectlink.com.np — built and deployed by a 2-person team.

The Problem

A local gym was managing member attendance, subscriptions, and payments entirely on paper and spreadsheets. Staff were spending significant time on manual check-ins, and there was no visibility into who had paid, whose membership had expired, or daily inventory levels.

My Role

Solo freelance project — designed, built, and delivered the full system end-to-end for a paying client.

What I Built

A Django backend with separate apps for members, billing, attendance, inventory, and cafe management. Facial recognition attendance runs on InsightFace with liveness detection to prevent photo spoofing. Role-based authentication separates admin, staff, and member access levels. Payment recording includes an approval workflow with suspicious-activity flagging, and membership expiry automation keeps the system self-maintaining. Everything ties together through a centralized dashboard with full attendance history.

Key Technical Decision

Used InsightFace with ONNX runtime rather than a cloud face recognition API — kept all biometric data on the client's own server, which was a hard requirement for privacy and cost reasons. Added passive liveness detection to prevent photo spoofing attacks.

Outcome

Delivered to a paying client — handles 100+ daily facial check-ins.

The Problem

E-commerce platforms rely on recommendation engines to drive engagement, but most implementations hide behind library abstractions. I wanted to build one from scratch to understand the algorithm layer — not just call a function that returns similar products.

My Role

Solo project — designed and implemented the full ML pipeline from data preprocessing to recommendation output.

What I Built

Text preprocessing and tokenization of product descriptions, converted into numerical feature vectors using Bag of Words vectorization. Cosine similarity is computed between all product vectors to find the closest matches, then results are ranked by similarity score and returned as a top-5 recommendation list.

Key Technical Decision

Used raw Bag of Words over TF-IDF intentionally — to understand why frequency-only representation loses context. That understanding directly informed the semantic search approach I used in PerfectLink, where embedding-based search solved this exact limitation.

Outcome

Trained on 10,000 Kaggle products — produces relevant top-5 recommendations.

The Problem

Network intrusion detection systems typically rely on labeled attack data, which is expensive to collect and constantly outdated. I wanted to explore whether an unsupervised approach could detect anomalies without any attack labels — training only on what "normal" looks like.

My Role

Solo project — researched, implemented, and evaluated the full deep learning pipeline from data preparation to detection threshold tuning.

What I Built

An autoencoder that trains exclusively on normal network traffic, learning to compress and reconstruct benign patterns efficiently. When anomalous traffic flows through, the model produces a high reconstruction error — that error becomes the detection signal. A tuned threshold controls the tradeoff between false positives and missed threats, which is the core engineering challenge in any production anomaly detection system.

Key Technical Decision

The interesting challenge wasn't the model architecture — it was threshold selection. Too low and everything flags as an anomaly. Too high and real attacks slip through. This tradeoff is the core engineering problem in production anomaly detection systems.

Outcome

97–99% detection accuracy on NSL-KDD — trained on normal traffic only.