DEV Community: Ihsan Andrinal

From Industry to University: Building 'TariffIntellect' with Gemini & Antigravity

Ihsan Andrinal — Wed, 21 Jan 2026 02:54:38 +0000

This is a submission for the New Year, New You Portfolio Challenge Presented by Google AI

About Me

For 2026, I am embracing a leverage in my career: evolving from a trade policy expert into a Domain-Expert Developer.

Currently, I am a Master of Information Technology student researching the intersection of Artificial Intelligence and International Trade Law. With over 10 years of professional experience in the trade facilitation industry, I realized that policy expertise alone isn't enough to modernize global trade.

My unique style is "Pragmatic Innovation." I don't just build code for the sake of coding; I build code to solve specific, high-stakes legal problems. My portfolio expresses my goal to bridge the gap between legal policy and modern technology using Neuro-Symbolic AI.

Portfolio

Here is TariffIntellect, a "Human-in-the-Loop" AI prototype deployed on Google Cloud Run. It demonstrates how AI can be used safely in regulated industries by combining visual recognition with a strict audit trail.

How I Built It

To build this prototype in record time (under 24 hours!), I leveraged the full power of Google's modern development ecosystem.

The Tech Stack:

AI Model: Google Gemini 2.0 Flash (via google-generativeai SDK).
IDE: Google Antigravity (Project IDX).
Backend/Frontend: Python 3.11 & Streamlit.
Infrastructure: Google Cloud Run (Dockerized).

🧠 The "Smart Routing" Logic

One feature I didn't want to compromise on was flexibility. The app follows a Smart Routing Architecture similar to real-world Customs lanes:

The Green Lane (AI Autopilot):
If a user uploads a general item (e.g., a Shoe), the system detects no conflict in the ambiguity_map.json. In this case, Gemini 2.0 Flash automatically analyzes the image and generates the HS Code using its vast internal knowledge base.
The Red Lane (Human Intervention):
If the system detects a high-risk keyword (like "Knife" or "Drone"), it triggers the Ambiguity Engine. The AI pauses, presents the specific legal conflict to the user, and waits for human input before finalizing the classification.

Development Process:

Rapid Scaffolding with Antigravity: I used Google's AI-first IDE to generate the initial project structure. By prompting the agent to "Create a Streamlit app structure with an ambiguity_map.json for logic handling," I skipped minutes of boilerplate setup.
Connecting the "Eyes": I integrated Gemini 2.0 Flash to act as the visual sensor. It identifies objects rapidly before passing the data to my Python logic.
Deploying to Cloud Run: I used the gcloud CLI to deploy the container directly from my terminal, ensuring the app is scalable and publicly accessible.

# My deployment command used for this challenge
gcloud run deploy tariff-intellect \
  --source . \
  --region us-central1 \
  --allow-unauthenticated \
  --labels dev-tutorial=devnewyear2026

What I'm Most Proud Of

I am most proud of the "Ambiguity Engine" architecture.

In the customs industry, "Black Box" AI is dangerous. A simple misclassification (e.g., mistaking a combat knife for a kitchen utensil) can lead to duty evasion and national security risks. Relying solely on probability scores is not enough.

I am proud that I successfully built a Hybrid Neuro-Symbolic System that solves this:

It Detects Legal Conflicts: Instead of blindly trusting the AI's vision, my Python logic intercepts the result and cross-references it with a deterministic rule set (ambiguity_map.json).
It Enforces Auditability: When a high-risk item is detected, the system forces a "Human-in-the-Loop" intervention. This creates a transparent decision trail that is legally defensible.

🚧 Lesson Learned & Future Roadmap

While this prototype successfully demonstrates the "Human-in-the-Loop" concept, a production-ready version would require:

Dynamic Rule Database: Currently, the Ambiguity Engine uses a static JSON file for demonstration purposes. In a real-world deployment, this would connect to a live SQL database of National Customs Rulings.
Confidence Thresholds: For the "Green Lane" (AI Autopilot), future versions will include a confidence score check. If Gemini is less than 90% sure, it should automatically route the item to the "Red Lane" for human review, reducing the risk of hallucinations.

This project proves that with the right combination of tools (Gemini for perception + Python for logic), we can build AI that is not only smart but also compliant, safe, and trustworthy to facilitate global trade.

First Submission Fuels Growth

Ihsan Andrinal — Fri, 01 Aug 2025 14:40:33 +0000

This is a submission for the World's Largest Hackathon Writing Challenge: After the Hack.

What's Next for Me and My Project
This marks my first challenge submission with the Algolia MCP Server Challenge, where I developed the HS Code Commodity Search Engine, integrating Algolia and Claude AI for tariff analysis. Not long after submitting, I received an exciting reach-out from the founder of Tarivio AI to collaborate and enhance his product’s workflow with my project. This partnership holds great promise for expanding the tool’s reach, potentially into Indonesia and beyond, and I’m considering open-sourcing more features on GitHub to foster community growth.

Reflect on What I Learned
This experience, my first in a coding challenge, taught me valuable lessons. I mastered securing API keys with environment variables and managing Git history after an initial exposure, while deploying to Heroku and Vercel sharpened my DevOps skills. More importantly, it inspired me to streamline this project with my current trade and logistics work, aiming to deliver impactful solutions for stakeholders like exporters and government agencies through efficient tariff management and compliance. This month shifted my focus from competition to creating meaningful value, shaping my trajectory toward innovative trade technology.

I’d welcome your insights or suggestions in the comments!

Building an HS Code Commodity Search Engine with Algolia and Claude

Ihsan Andrinal — Mon, 28 Jul 2025 03:40:26 +0000

This is a submission for the Algolia MCP Server Challenge

What I Built

I developed an HS Code Commodity Search Engine for the Algolia MCP Server Challenge. This web app enables users to search Harmonized System (HS) codes, filter by categories and tariffs, and analyze tariff differences using Anthropic’s Claude AI. Hosted with a Node.js backend on Heroku and a React frontend on Vercel, it targets trade-related needs.

Benefit for Users:

For exporters, tariff analysis offers critical insights before exporting, aiding decision-making.
For government users, Algolia provides a searchable, public-facing tariff database, simplifying commodity information access.

Key Project Features

Search: Algolia powers fast HS code and commodity lookups.
Filters: Category and tariff filters refine results for a tailored experience.
Tariff Analysis: Claude AI delivers real-time tariff difference insights.

Demo

Explore the live demo at Vercel URL and YouTube. Search "horses," filter by "Live Animals," and click “Analyze Tariffs with Claude” to see the analysis.

Repo

The code is on GitHub under the MIT License, open for collaboration.

How I Utilized the Algolia MCP Server

I used the Algolia MCP Server to index a dataset of 39 dummy commodities with HS codes, tariffs for Indonesia and the USA, and categories. The backend leverages Algolia’s search API for efficient queries and facet-based filtering, demonstrating robust platform utilization.

Innovative Approach to Algolia Usage

The innovation combines Algolia’s search with Claude AI for tariff analysis. While Algolia retrieves data, Claude processes it for insights, enhanced by planned synonym support (e.g., "timber," "lumber" for "wood") to boost relevance.

How It Worked Step by Step

Data Preparation: Parsed a CSV with dummy HS code data.
Algolia Indexing: Indexed data with facets for filters.
Backend Setup: Built a Node.js server with Express for search/enrich endpoints.
AI Integration: Integrated Claude API for analysis.
Frontend Development: Developed a React app with Tailwind CSS, linking to the backend.
Deployment: Deployed on Heroku and Vercel, ensuring port compatibility.

Key Challenges

CSV Parsing Issues: Handled malformed dummy data with validation.
Duplicate Records: Added deduplication logic in Algolia.
Facet Configuration: Iterated to ensure accurate filtering.

Key Takeaways

Algolia’s speed and facets transform data-driven apps.
Claude integration adds value but requires API optimization.
Heroku/Vercel deployment was streamlined, though port setup was tricky.

Next Development and Current Limitations

Current Limitations: The app uses dummy data comparing Indonesia and USA tariffs, covering only 39 commodities—not all HS codes.
Next Steps: Expand to other countries with multilingual support (e.g., Spanish, Mandarin), automate tariff updates with n8n, and configure Algolia to recognize synonyms like "timber" and "lumber" as related to "wood" for better relevance.

Conclusion

Thanks to the DEV Community and Algolia MCP Server Challenge team! The project is open-source under the MIT License—contribute on GitHub. Test it at Vercel URL!