Figarri Keisha

Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training

Mon, 01 Sep 2025 00:00:00 +0000

All for law and law for all: Adaptive RAG Pipeline for Legal Research

Fri, 01 Aug 2025 00:00:00 +0000

🛒 Building a WhatsApp-Based Grocery Assistant in 3 Days Using n8n

Thu, 10 Jul 2025 00:00:00 +0000

1. Background

As a student living in the UK, balancing study, life, and cooking has been a challenge. Between coursework, lectures, and late-night study sessions, I often forget what’s already sitting in my fridge. This leads to the all-too-familiar outcome: buying groceries I already have, letting things expire, and ultimately wasting food and money.

I wanted a solution that didn’t require me to change my habits or download new apps, just something that worked where I already spend time: WhatsApp.

2. The Solution

I built a Grocery Assistant, an AI-powered tool that helps me:

Extract items from receipt images and store them in a live inventory table
Remind me of expiring items and help generate grocery lists
Communicate entirely through WhatsApp, which I already use daily

This means no extra app installations, no switching platforms, just smart grocery help through a familiar interface.

🧪 Demo

Below are three short demo clips that showcase the main features of the WhatsApp-based Grocery Assistant:

📸 1. Receipt Upload → Inventory Update

A user sends a photo of a grocery receipt via WhatsApp. The assistant:

Detects it as a receipt
Extracts the items using AI
Adds them to the inventory table
Responds with a bullet-point confirmation of what was added

🛒 2. Generate Grocery List from Expiring Items

The user asks, “Give me grocery list based on expiring items”

The assistant checks for items that are low in quantity or expiring soon and replies with a smart grocery list.

❌ 3. Upload Non-Receipt Image

A user sends a random photo (e.g. a street sign or a product box). The assistant:

Classifies it as not a receipt
Transcribes visible text (if any)
Responds politely, explaining that no grocery extraction was possible

🛢️ Using PostgreSQL with Supabase

To store and manage the grocery data, I used PostgreSQL hosted on Supabase. The schema consists of two key tables:

1. `product_shelf_life`

This is a reference table containing general knowledge about how long common grocery categories typically last. For example, Milk expires in 12 days, Vegetables in 3 days, Snacks in 180 days, etc.

2. `inventory`

This table records items extracted from receipts and keeps track of:

Item name
Product category
Quantity and unit
Inferred expiry_date
Timestamp of purchase

The expiry_date is automatically calculated as:

expiry_date = current_date + expected_expired_days

3. Prototyping with n8n

To build this fast, I used n8n, an open-source workflow automation tool. Why n8n?

✅ Low-code with visual logic
✅ Easy integration with WhatsApp (via Business API)
✅ Supports AI models and custom databases
✅ Fast iteration; I built and tested this in under 3 days

For prototypes like this, n8n is perfect. It’s modular, visual, and allows integrating AI tools with databases, HTTP endpoints, and messaging platforms with minimal boilerplate.

4. Workflow Breakdown

Let me walk you through the system design, as shown in the visual below:

🟫 Grocery Assistant

When a user sends a text message, the Grocery Assistant (powered by Gemini 2.5 Flash) processes it. It chooses the right tool based on intent: whether it’s to create a grocery list, check for expiring items, or ask about what’s in stock.

✅ Tools used:

Postgres Chat Memory to access inventory data
Expiring Items to check for near-expiry products (connected with PostgreSQL)
Replies are sent directly back via WhatsApp.

🟦 Image Classifier

If the input is an image (e.g., a photo of a receipt), it’s routed to the Image Classifier. This step uses Gemini 1.5 Flash to classify whether the image is a receipt or something else.

📌 Outcome:

If it’s a receipt → go to receipt extractor.
If not → transcribe and send the text back with a polite fallback.

🟪 No Receipt Response

If the image is not a receipt, the system still makes use of it. The agent transcribes visible text using Gemini 1.5 Flash and replies via WhatsApp with the message:

“This image does not appear to be a receipt, so no extraction was performed.”

This ensures a graceful fallback and lets users know what’s happening.

🟩 Smart Receipt Extractor

When a valid receipt is detected, the image goes through the Smart Receipt Extractor workflow:

🔗 Merge the binary image
📄 Extract raw text using Gemini 2.5 Flash
🗂️ Match products to known categories via database
🧠 Parse structured items with Gemini 1.5 Flash
🧹 Clean and transform the data (add expiry based on category shelf life)
🔄 Split items for individual inserts
🛒 Insert into PostgreSQL inventory
📬 Aggregate and send a WhatsApp checklist of added items

Despite being entirely no-code, this workflow handles multi-model orchestration, structured output parsing, and real-time database updates.

5. Conclusion

The Grocery Assistant works and it already saves me time and reduces food waste. It’s still a prototype, but the foundation is solid and extensible.

🛠️ What can be improved?

🔔 Add a cron job to send WhatsApp reminders when items are close to expiry
🧠 Improve product mapping by adding more granular shelf-life estimates per product name (not just by category)
🤖 Better handling of OCR errors from low-quality receipts
📊 Build a small dashboard to visualize inventory freshness trends

🧠 Final Thoughts on n8n

While I haven’t yet explored the feasibility of public, large-scale deployment using n8n, I found it incredibly powerful for prototyping and rapid iteration. Within just a few days, I was able to build a fully functional AI-powered grocery assistant with real database integration and WhatsApp automation, all without writing a full backend from scratch. For now, I see n8n as a reliable tool for internal workflows, testing ideas quickly, or building MVPs.

👉 Download the n8n workflow of the Grocery Assistant here

An Automatic Document Validation System in the Platform Merdeka Mengajar by Using Word Embeddings and Supervised Classification Models

Wed, 01 Jan 2025 00:00:00 +0000

You can read the high-level explanation through this medium.

🎓 The GovTech Edu Data Science Team’s Role in Transforming Education in Indonesia

Thu, 18 Apr 2024 00:00:00 +0000

Overview

This story, specially crafted by Andika Rachman Hakim, Senior Data Scientist and my direct team lead, offers a behind-the-scenes look at how our compact Data Science team at GovTech Edu addressed some of Indonesia’s most pressing educational challenges through AI and data innovation. From automating the validation of 500k+ teacher submissions per month, developing recommendation systems, to launching a GenAI-powered metadata tool. The article highlights our journey building impactful, scalable public-sector solutions through tight collaboration and principled product thinking.

👉 Read the full story on Medium

🌾 Empowering Local Farmers with Data - KaPaN

Mon, 06 Nov 2023 00:00:00 +0000

Extreme weather is inevitable, but crop failure doesn’t have to be.

🏆 Regional Recognition at the United Nations Datathon 2023

As Team 4SKA1, we were honored to be selected as The Best Team in the Asian Region at the United Nations Datathon 2023. This global competition challenged participants to develop innovative data-driven applications combining geospatial and statistical data to advance the Sustainable Development Goals (SDGs).

🚜 The Problem: Climate Uncertainty Meets Agriculture

Farmers in West Java, one of Indonesia’s major rice-producing regions, are increasingly vulnerable to extreme weather events, such as droughts, floods, and irregular rainfall. These threats can severely disrupt food supply and jeopardize farmers’ livelihoods.

Despite the urgency, many farmers lack access to early and localized information to prepare for such risks.

💡 The Solution: KaPaN

We proposed KaPaN (Kalender Padi Nusantara - Nusantara Rice Calendar), a prototype website that serves as an early warning system for extreme weather, tailored specifically for local farmers.

KaPaN empowers users to:

Track when and how long extreme weather is forecasted in their area
Visualize the spatial extent of weather anomalies
Report and share real-time ground-level information with fellow farmers

The name “KaPaN” (Indonesian for “When?”) reflects our goal to answer farmers’ biggest question when facing uncertainty.

🔬 The Method: Extreme Value Analysis

We used Extreme Value Analysis (EVA) (specifically the Block Maxima method) to analyze rare and high-impact events such as intense rainfall and drought.

With the help of Python and the pyextremes library, we simulated thresholds for extreme conditions and provided early warnings that farmers can act on.

🌐 Accessible, Actionable, and Community-Driven

The KaPaN platform is designed to be:

User-friendly: intuitive dashboard tailored for farmers
Insightful: visualizations help interpret trends over time and across regions
Collaborative: enables farmers to contribute ground-level observations

📂 Learn More

🤝 Meet the Team – 4SKA1

We are a passionate group of data practitioners from Indonesia who believe in the power of data for local impact:

Bagoes Rahmat Widiarso
Bhaskoro Muthohar
Figarri Keisha
Muhammad Nassirudin

Experience

Tue, 24 Oct 2023 00:00:00 +0000

🍚 Food Insecurity Early Warning System

Fri, 11 Nov 2022 00:00:00 +0000

Overview

At the 2022 United Nations Big Data Hackathon, Team 4SKA1, formed by young Indonesian data scientists including myself, took on the global challenge of Food Insecurity. Our solution was a scalable Early Warning System prototype built using open data and visualized via Google Data Studio. The project won Best Visualization Award in the Youth Track for its clarity, insight, and practical utility.

The Challenge

Climate change and socio-economic instability are increasingly affecting global food security. Yet, early warning systems are often absent, especially in vulnerable regions. Our task was to answer:

Can we monitor and anticipate food insecurity trends globally?
How does a country’s food insecurity affect others through trade flows?

Our Approach

We created a predictive model using data from:

The UN Hackathon 2022 datasets
World Bank, AtlasBig, and OurWorldInData

We engineered three key scores for each country:

Socio-Economic Score (SES): clustering countries based on unemployment, undernourishment, mortality, etc.
Climate Score: using PCA to combine metrics like CO₂ emissions, energy use, forest area.
Food Score: integrating food production, consumption, waste, and trade patterns.

These were combined to train a linear regression model to estimate national food insecurity scores with ~10% error.

The Dashboard Prototype

We built an interactive Google Data Studio dashboard to visualize:

Global food insecurity status by country
SES, Climate, and Food score comparisons
Country-level food trade relationships
Exploratory visualizations to empower policy and humanitarian decisions

👉 View the Dashboard

📄 Download Full Project PDF

📈 Causal Inference - Can We Trust What Happens After?

Fri, 25 Mar 2022 00:00:00 +0000

Overview

How do you know if your new business strategy actually worked or if the trend was going to happen anyway? In this article, I introduce the Difference-in-Differences (DiD) method, a powerful statistical technique for estimating causal impact from observational data, using a real case study from my work at Bukalapak.

We walk through the full pipeline: from identifying the problem with naive before-after comparisons, validating assumptions like parallel trends, to implementing DiD with Python regression models and even conducting a placebo test. I also showcase a custom DiD Python library that makes the process seamless and reproducible.

Read the full article on Medium and discover how we turn business questions into causal answers.

🛍️ Predicting Users’ Intention When Searching on Online Marketplace Platforms

Tue, 25 Feb 2020 00:00:00 +0000

Overview

Ever wondered why some people search with laser focus, while others just explore? In this post, I walk through how we tackled this challenge at Bukalapak by modeling users’ search intentions, specifically whether they are exploring or locating. Using historical behavioral data, we built an unsupervised learning model to predict these intent types based on queries, enabling personalized search strategies and a better user experience. Curious how a simple keyword like “iPhone 5” can lead to very different behaviors? Dive into the full story to find out.

👉 Read more on Medium

Projects

Mon, 01 Jan 0001 00:00:00 +0000

Figarri Keisha

Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training

All for law and law for all: Adaptive RAG Pipeline for Legal Research

🛒 Building a WhatsApp-Based Grocery Assistant in 3 Days Using n8n

1. Background

2. The Solution

🧪 Demo

📸 1. Receipt Upload → Inventory Update

🛒 2. Generate Grocery List from Expiring Items

❌ 3. Upload Non-Receipt Image

🛢️ Using PostgreSQL with Supabase

1. product_shelf_life

2. inventory

3. Prototyping with n8n

4. Workflow Breakdown

🟫 Grocery Assistant

🟦 Image Classifier

🟪 No Receipt Response

🟩 Smart Receipt Extractor

5. Conclusion

🛠️ What can be improved?

🧠 Final Thoughts on n8n

Did you find this page helpful? Consider sharing it 🙌

An Automatic Document Validation System in the Platform Merdeka Mengajar by Using Word Embeddings and Supervised Classification Models

🎓 The GovTech Edu Data Science Team’s Role in Transforming Education in Indonesia

Overview

Did you find this page helpful? Consider sharing it 🙌

🌾 Empowering Local Farmers with Data - KaPaN

🏆 Regional Recognition at the United Nations Datathon 2023

🚜 The Problem: Climate Uncertainty Meets Agriculture

💡 The Solution: KaPaN

🔬 The Method: Extreme Value Analysis

🌐 Accessible, Actionable, and Community-Driven

📂 Learn More

🤝 Meet the Team – 4SKA1

Did you find this page helpful? Consider sharing it 🙌

Experience

🍚 Food Insecurity Early Warning System

Overview

The Challenge

Our Approach

The Dashboard Prototype

Did you find this page helpful? Consider sharing it 🙌

📈 Causal Inference - Can We Trust What Happens After?

Overview

Did you find this page helpful? Consider sharing it 🙌

🛍️ Predicting Users’ Intention When Searching on Online Marketplace Platforms

Overview

Did you find this page helpful? Consider sharing it 🙌

Projects

1. `product_shelf_life`

2. `inventory`