Starts from raw, noisy sensor streams: cleaning, feature engineering, and exploratory analysis — heatmaps and figures — to identify which signals actually predict a coming fault, before any of it reaches a model. Forecasts machine state and flags anomalies 3–4 days before they happen. A locally-hosted LLM, fine-tuned on internal domain knowledge, turns the forecast into a plain-language read that data analysts can act on immediately. Used daily for fault analysis across multiple equipment types, and has held up reliably in production.
An Elo-based match model with Monte Carlo simulation — a real group stage, three-way win/draw/loss outcomes, and a live bracket that simulates forward from FIFA's actual knockout stage. Called the Final for Spain at 57.8%, validated against the real result (Brier score 0.178).
View on GitHub ↗Real-time shelf monitoring with Raspberry Pi and OpenCV — auto-detects low stock and sends LINE Notify alerts. This project is what led directly to my current role.
★ Pre-Capstone Exhibition AwardAn automated opening/closing mechanism for 15–50cl screw-type lab containers, built with SolidWorks CAD, Arduino, servo motors, IR sensors, and electromagnets. Tested for cycle time and reliability: under 3 seconds in 90%+ of trials, with an overall success rate above 90%.
★ Capstone Exhibition Award 2025Before I wrote a line of Python, I spent two and a half years hosting a podcast — not because I wanted to be on air, but because I was obsessed with understanding people: why they feel what they feel, what they're really saying beneath the surface. That obsession didn't disappear when I switched to engineering. It just found a new medium.
At フクシマガリレイ株式会社, that's the discipline behind every pipeline I ship: observe first, understand what the data is actually saying, then build something a non-technical colleague can rely on without ever opening Python.
Building ML models, running time-series and data analysis, and developing the full software stack — FastAPI, Streamlit, MySQL — that delivers AI to the people who need it.
Production engineering internship focused on BOM management, routing optimization, and interplant communication — used data-driven problem solving to improve production data quality.
Hosted a Nepali podcast for 2.5 years. Built real skills in communication, storytelling, and understanding audiences — skills that turned out to be surprisingly relevant in engineering.
During the 2020 lockdown, youth in the neighborhood had lost touch with each other. Co-founded a club to rebuild it: gathered 30 members from the municipality and started running sanitation, awareness, and fellowship programs for local youth — including a sanitation initiative at Champadevi Temple, one of several the club has run since.
Most ML models output a prediction with misplaced confidence. Reliable AI — systems that quantify their own uncertainty and defer appropriately — is one of the most underexplored problems in applied ML. I'm currently exploring conformal prediction and calibration as practical paths toward it.
Open to conversations about ML systems, industrial AI, research collaborations, or just an interesting problem you haven't solved yet.
Based in Osaka · Open to work · Open to remote collaboration