Rafi Aqila Hidayat

1st Place, Synogize AISQL Innovation Challenge

AISQL Site Safety Inspection

AI-powered site safety inspection app built entirely on Snowflake. Upload a photo from a construction site and get a hazard report, a risk score, and corrective action recommendations in seconds.

How it works

• The app takes a photo from a construction site and runs it through three Snowflake Cortex AI functions
• AI_FILTER checks whether the image actually shows a hazard worth flagging
• AI_CLASSIFY sorts the hazard into a category: missing PPE, unsafe scaffolding, blocked exits, and so on
• AI_COMPLETE writes a plain language explanation of the risk and suggests what to do about it
• Each site gets a running risk score out of 10, built from the history of flagged images at that location
• The dashboard shows which sites are trending worse over time, not just a single snapshot
• When a site crosses a risk threshold, the system sends an email automatically so the safety team finds out the same day instead of at a weekly review
• Users can export a full report as HTML or pull a checklist as CSV for site audits

Tech stack

Personal Project

Tennis Tracker

I play tennis two to three times a week and had no real way to tell if I was actually improving or just having good and bad days. So I built a system to track it properly.

Logging a session

• Typing /log on Telegram starts a short conversation
• Asks about session type, location, skill ratings (serve, forehand, backhand, volley, footwork, focus, return), energy before and after, mood, discomfort, match result, and free text notes
• About 15 data points per session, all saved to BigQuery

AI coaching, three ways

• After every session, Gemini sends back a short coaching summary based on what was logged
• Every Sunday, a weekly recap covering skill trends, energy patterns, and a plan for the week ahead
• The /insights command pulls a 90 day report on demand, with skill trends, win and loss analysis, and recurring patterns

Asking questions directly

• Free form questions like "what should I focus on this week" or "why have I been losing more matches lately"
• Answers using full session history
• Remembers earlier parts of the conversation, so follow up questions do not need repeated context

Scheduling in plain English

• A message like "schedule match at Titiwangsa tomorrow 7pm against Fara" gets parsed automatically and added to Google Calendar
• The night before, a pre session brief gets sent with focus areas, a quick game plan, notes from the last few sessions, and the weather forecast

The dashboard

• A Telegram Mini App with four tabs
• Overview: session counts, win and loss record, skill radar chart
• Progress: skill trends over time with AI generated notes on what is improving and what needs work
• Matches: win rate by opponent level, head to head records
• Activity: 30 day heatmap of session frequency

Tech stack

AI Chatbot & Operations Automation

This project is built with another developer in two phases for an agribusiness company in Indonesia that makes organic fertiliser and sells to business, government, and increasingly consumer customers across several countries.

Phase 1: the customer service chatbot

• Before this, field staff sent visit reports over WhatsApp that someone had to copy into a spreadsheet by hand, and there was no automated way to greet and qualify new customers before handing them to sales
• The chatbot answers product questions using a knowledge base built from the company's own documents
• It greets new customers and figures out what they need before passing qualified leads to the sales team through Telegram
• The knowledge base started as a system prompt built directly from company documents, with a fallback plan to move to RAG if document volume grew past what a system prompt could handle well
• The bot remembers returning customers by their WhatsApp number using a lightweight memory layer, so it does not treat every conversation as the first one

Phase 2: internal operations

• This phase moved from customer facing to internal operations, covering employee attendance, all production and inventory reporting, incoming order tracking, and the monitoring dashboards
• Attendance runs through WhatsApp with photo verification and live location, so HR can see who is on site without a separate form
• Production and inventory covers raw materials coming in, materials being used, finished goods coming out, and stock leaving the warehouse, each tracked separately so stock levels update automatically
• Incoming orders get an automatically generated order ID the moment they come in
• Order status updates on its own as production reports reference that ID, moving from new order to in production to produced without anyone touching a spreadsheet
• Three separate dashboards sit on top of all this: one for HR to see attendance, one for the production floor built for a TV display, and one for management with a full view across the operation

Tech stack

Undergraduate Thesis, Infineon Technologies

Semiconductor Capacity Optimisation

Undergraduate thesis on optimising capacity allocation in semiconductor manufacturing using Integer Linear Programming, developed in collaboration with Infineon Technologies and validated against real production data.

Key features

• Integer Linear Programming model for allocating capacity across photolithography machines
• Implemented and solved in MATLAB
• Validated with real production data from Infineon’s Kulim fab
• Built on time-study findings that the longest activity accounts for 42% of total process time

Tech stack