1
AetherDrop
secure file-sharingSecure file-sharing platform with expiring links, download limits and cloud-backed storage. ZIP streaming for multi-file downloads, fronted by a reverse proxy and tunnel.
Embedder
cold
Full-stack developer for production-grade systems and AI applications. This page is a working retrieval-augmented system over my career — it embeds, indexes and answers. Scroll to travel the pipeline, or query it live below.
Every career is a document set. Snippets peel off the pages and get pulled into the ingester — and that resume.pdf? Click it.
The parser walks every page and pulls clean text. Keep scrolling — the aperture opens and we go in.
Pages are scanned into tokens; the blades group tokens into overlapping chunks — structure stays whole.
Meaning becomes geometry — each chunk is projected into a space where "similar" literally means "close."
Every vector is wired to its nearest neighbours, so queries hop the map in logarithmic time.
Hybrid search casts a wide net; a reranker re-scores the shortlist. Three chunks survive.
Answer using ONLY the retrieved passages. Cite as [n]. Refuse if uncovered.
What has Nipun shipped to production?
Written against the passages only, every claim cited — with a polite refusal when similarity runs dry.
The pipeline above is a schematic on purpose. These are the full-size units — end-to-end ownership, from idea to working infrastructure.
Secure file-sharing platform with expiring links, download limits and cloud-backed storage. ZIP streaming for multi-file downloads, fronted by a reverse proxy and tunnel.
AI system predicting hospital readmissions from clinical notes. Fine-tuned ClinicalBERT on MIMIC-III behind a Flask inference API, with a clinician-facing dashboard on a microservice architecture.
Offline AI assistant for emotional support in constrained environments. Local LLaMA inference via Ollama, emotion detection with DeepFace + OpenCV, and a sentiment-aware conversational system with session tracking.
NLP-based fake-news detection engine with explainability. BERT trained on 72,000+ articles, confidence scoring, and an API-ready classification engine.
ML system predicting sustainability scores from lifestyle data. Regression and ensemble models with careful feature engineering and a results dashboard.
Blockchain-based system for tracking ownership of cultural assets. Solidity smart contracts power an immutable ownership ledger with fraud-resistant provenance tracking.
On first query it loads a neural embedding model (all-MiniLM-L6-v2) entirely in your browser, embeds my career into 384-d vectors, and retrieves by true cosine similarity. It understands intent — so ask simple things ("what's your name?") or specific things ("tell me about AetherDrop") and it answers like a person, not a spec sheet.
The index is cold. Ask something — it warms (loads the model) on first query.
—