AI Engineer | Software Developer
I build AI-powered software, from research prototypes to production services. My work spans knowledge-graph retrieval, real-time voice AI, autonomous agents, and the backend services and web interfaces around them.
- Currently: AI Automation R&D Intern at Advfiction, designing workflow automation for a multi-sector Business OS.
- Recognition: Two-time Smart India Hackathon internal qualifier (2025 and 2026), selected among the top 45 of 300+ team entries.
| When | Role | What I did |
|---|---|---|
| Aug 2026 to Present | AI Automation R&D Intern, Advfiction (remote) | Designing automation workflows for a multi-sector Business OS that connects CRM, ERP, AI and external services. Evaluated n8n, Zoho Flow, Zapier and Workato for webhooks, scheduled jobs, conditional logic, retries and error handling. |
| May to Jul 2026 | Research Intern, Cyber Threat Intelligence | Built a GraphRAG framework (ChromaDB, SentenceTransformers, Neo4j) that converts unstructured threat reports into MITRE ATT&CK knowledge graphs for next-technique prediction. Optimized a RotatE embedding model in PyTorch that outperformed GNN, LLM and Markov-chain baselines. |
| Mar to Aug 2026 | Exchange Student, Universiti Teknologi Malaysia | Semester exchange program in Johor Bahru, Malaysia. |
| Jan to Mar 2026 | AI Engineering Intern, Simsy AI (remote) | Deployed production-ready voice AI pipeline components (Python, FastAPI, Docker, LightRAG, LiveKit) and cut audio payload latency by 32% on AWS. Optimized the asynchronous backend and REST API architecture to support more parallel users, and reduced container baseline storage by 45%. |
| Ongoing | Member, McCarthy Lab, Next Tech Lab | Active member of a student mentorship lab, collaborating on AI projects. |
A temporal-causal GraphRAG system that predicts the next technique an attacker group is likely to use, based on actor-aware knowledge graphs and RotatE embeddings.
- Result: 41.38% Hits@1
- Stack: Python, PyTorch, PyKEEN, Neo4j, ChromaDB
Nexi: AI Voice Assistant
A voice assistant that routes each question between campus PDFs and structured JSON data, with persistent conversation memory and RAG-based grounding to reduce hallucinations.
- Result: 340 ms response latency and 98.4% grounded response accuracy
- Stack: Python, LlamaIndex, ChromaDB, LiveKit, Gemini
An edge-based payment authentication system running on a Raspberry Pi 5. It uses near-infrared imaging, MediaPipe hand landmarks, Gabor wavelet VeinCode generation and parallel matching, with a FastAPI backend and a React app for enrollment, scanning and payment authorization. VeinCodes are stored encrypted.
- Result: Under 500 ms authentication and a false acceptance rate below 0.1%
- Stack: Python, MediaPipe, OpenCV, FastAPI, React, SQLite, Raspberry Pi 5
A LangGraph agent that parses payment failures, classifies the root cause and sends personalized retry links, reducing manual intervention. Event-driven FastAPI webhooks let it run as middleware inside existing payment backends without changing client infrastructure.
- Stack: Python, LangGraph, FastAPI, SQLite
Gold marks the tools I use most.