Sebin Shaiju

Projects

Nine builds, ordered newest first. The ones with a metric attached are the ones where I measured something; the rest are honest about being unmeasured.

HIRE4GIGpatent filed
Income guarantees for gig workers that adapt every three minutes, matched against orders with a fairness-aware cost function.
Python · Gaussian process regression · Hungarian algorithm
82% ↓ cost
2026
Your.Store
A Shopify operations copilot where the model plans and deterministic Python executes. Every write is gated behind a human approval it cannot route around.
LangGraph · FastAPI · MCP · AGPL-3.0
18 nodes
2026
Electronics storefront, Nepalclient
Full-stack commerce platform for the Nepali market: NPR pricing, connectIPS payments in progress, typo-tolerant Postgres search, and an admin panel scoped to the website rather than the business.
Next.js 16 · TypeScript · Prisma · Supabase
in build
2026
render-langgraph
Renders a LangGraph topology in the browser with no code changes and no LangSmith account. Published on PyPI as a v0.1 walking skeleton.
Python · PyPI
v0.1
2026
myOS Real Estate
Hackathon prototype from Cursor × eVoost at Hub71. A FastAPI intelligence microservice for real estate: deterministic property and district rankings, semantic search over OpenStreetMap amenity data, and an LLM layer that only explains and routes rather than computes.
FastAPI · LangChain · vector search over OpenStreetMap
1 day
2026
Subject-leakage auditorin review
An audit tool for EEG pipelines that catches the most common way a brain-decoding result is accidentally inflated: the same subject appearing on both sides of the split.
Python · EEG · scikit-learn
paper
2026
Pneumonia detection ensemble
Three CNNs averaged over 5,863 chest X-rays, with Grad-CAM heatmaps so a radiologist can see what the model looked at.
TensorFlow · DenseNet121 · ResNet50 · MobileNetV2
88.8% acc
2025
This site
Static, markdown-backed, edited from the browser. No database, because nothing here changes without me.
Astro · Netlify
n/a
2025
Do LLMs take a wasteful path to their answers?
An interpretability study asking whether a transformer's internal representation wanders unnecessarily on the way to an answer, and whether that slack could be exploited for cheaper inference. It couldn't: the wandering turned out to be load-bearing.
Python · Qwen2.5 · logit lens · cosine similarity
9× wandering
2026