Projects

Production portfolio with Avocado AI, a streaming RAG agentic chatbot backed by a 4-stage hybrid retrieval pipeline: query expansion (up to 4 variants), batched dense search via ChromaDB (all-MiniLM-L6-v2 ONNX), BM25 lexical search (rank_bm25), and Reciprocal Rank Fusion (k=60) — all before Gemini 2.5 Flash sees the question. Implement model fallback mechanism via Gemini, Grow and OpenRouter with the knowledge base of ~80 atomic documents auto-syncs incrementally: new or edited documents are upserted, deleted documents are purged from ChromaDB, unchanged ones are skipped entirely. Runs as a FastAPI Docker container on AWS Lightsail with zero-downtime blue-green deployments, Nginx + Let's Encrypt HTTPS, daily S3 backup of SQLite analytics, and a Next.js 16 static frontend on GitHub Pages.

Achieved 15ms token latency on Snapdragon NPUs — a 10× improvement over cloud inference — by fine-tuning Llama 3.2 3B on security logs with QLoRA and deploying via 4-bit AWQ quantization through ONNX Runtime on-device. Guaranteed zero data loss during network partitions via an offline-first SQLite buffer with background sync workers.

AI-powered genomics discovery platform currently being built at NYU College of Arts and Science. Focused on making genomic data exploration accessible and actionable for researchers.

In active development at NYU CAS.

Production LangGraph + Llama 3.1 70B system that semantically maps global researcher collaboration networks by indexing millions of papers from Elsevier's Science Direct/Scopus. Cut P99 RAG latency by 78% with Write-Through Redis caching; sustained 99.9% uptime at 3,000+ RPS on AWS ECS.

Conflict-free simultaneous multi-user editing using Yjs (CRDTs) and WebSockets, scaled horizontally via Nginx load balancing across containerized instances. Increased AI auto-complete context quality by 65% with a Context-Aware Coding Agent using AST-based chunking and Voyage-Code-2 embeddings.

High-traffic progressive web app for academic grade forecasting at VIT. Scaled to 17K+ monthly active users and around 20K+ registered accounts. Achieved #2 Google Search ranking via programmatic SEO with sub-second mobile load times.

After six years in production and 17,000+ monthly users, gradeVITian has been rebuilt from the ground up — re-engineered on Next.js and FastAPI and relaunched at gradevitian.jayaremala.com. Heartfelt thanks to the VITian community (Class of 2020–2025) for making the original such a success; this journey defined me as an engineer, and this new chapter carries it forward.

A simple online web application that helped people find my dad's business in a mid-sized town in India.

At its heart, this was a project of purpose: an online gateway built to help people navigate my father's new business within our town. Seeing a simple web application drive real-world sales and provide stability for my family during the COVID-era was a transformative experience. It taught me that engineering is most powerful when it serves a personal and communal need.