I build backend systems, AI pipelines, and full-stack applications. Right now I'm working on semantic intelligence infrastructure, serverless document processing, and agentic AI workflows. I care about systems that are clean underneath and quietly reliable in production.
You can find me on GitHub, LinkedIn, LeetCode, or reach me over email.
Refactored a monolithic backend into a clean service-layered architecture, deployed self-hosted embedding infrastructure on GCP (mxbai-embed-large-v1, ~930ms latency, 80% cost reduction), and built a semantic intelligence pipeline for Shadow IT detection using cosine-similarity ranking and threat clustering. Orchestrated the full Docker Compose deployment behind Caddy with automated SSL.
Architected a Bronze/Silver/Gold medallion pipeline on AWS serverless converting Japanese insurance PDFs into structured, citation-grounded comparison tables with Google Gemini. Designed the JSON output schema, audited the backend for CORS, JWT, prompt injection and upload validation, and used parallel-agent patterns to speed up iteration.
A cyclic LangGraph workflow coordinating Chat and Profiling agents, with a ChromaDB RAG pipeline that filters product embeddings by active user constraints and a React frontend rendering AI-generated inventory and profile updates.
A Selenium automation engine (undetected-chromedriver, stealth, rotating proxies) for multi-account management, paired with a Groq/OpenAI/Gemini content pipeline and a Next.js dashboard with real-time session tracking, heatmaps, and analytics.
A database that goes beyond simple semantic similarity to build rich, interconnected knowledge representations — a superior mode of conceptual dwelling.
A full-stack application with extensive CRUD operations, API endpoints, and NoSQL database integration.
Thinking Within Images — on perception, movement, and what a system can be built to notice.
An embodied perceptual agent is not a model handed an image. It is a situated process with a locus, a limited body of organs, and a memory of what it encountered from there.
An image archive becomes a perceptual world only when what is there and what is noticed are allowed to remain different kinds of thing.
A percept is not a fluent sentence about an image. It is a durable act of noticing that can name what it rests on, how it was made, and where it can travel.
Tool calling becomes perceptual intelligence only when the tools are specialist organs with explicit limits, rather than decorative extensions of one model's voice.
The answer is neither a server nor a simulated 3D room. An agent lives at a locus in a perceptual topology, with a private history and access to a carefully curated shared world.
A system can move through images without inventing the continuous 3D world between them. It moves by following relations that instruments have earned.
The dominant architecture puts one large model at the centre and treats everything else as plumbing. There is a different shape available, and it is better suited to perception.
Between one image's axis of attention and another's, something can open. Treating that opening as a real object — not a turn of phrase — changes what a system can do.
Movement inside an image need not mean a camera travelling through reconstructed space. There is another kind — relational, perceptual — and it is the one that matters.
Tokenisation, SoftMax activation, and custom training loops.
Encoder, feature encoding, latent representation, sigmoid, decoder — no PyTorch.
Patch embedding, positional encoding, multi-head self-attention, QKV computation.
Text preprocessing, feature extraction, BoW vectorization.
750+ problems solved across 75+ contests, 55 of them rated.