TARA — TransJakarta AI Customer Service Chatbot | Arsyadam's Portfolio

Overview

TransJakarta serves millions of passengers every day, but many complaints still reach customer service incomplete: missing route numbers, card details, or a clear timeline of what happened. I built TARA, an AI customer service assistant that answers transit questions and guides passengers through official complaint reports across WhatsApp, Instagram, Facebook, and the TJ App.

Instead of one vague follow-up message, Tara walks users through a numbered checklist tailored to their complaint type, so reports reach the CS team with the information they actually need. You can try it today on Instagram @infotije, WhatsApp Customer Care Tara AI, or inside the TJ:TransJakarta app.

What Tara Can Do

Trips & routes

How to get from A to B, including fastest, cheapest, or fewest-transfer options
Check whether TransJakarta serves a given destination
Route info: which stops a corridor passes through
Operating hours and service schedules

Location & stops

Find the nearest bus stop from a place name or the user's location
Coordinates for a location or stop
Departure times at a specific stop (e.g. "What time does route 9 leave from Halte Pancoran Tugu?")

Service information

FAQs on TransJakarta: payment (Flazz, e-Money, QRIS, etc.), tickets, policies, tips, accessibility
Stop facilities: toilets, lifts, escalators, prayer rooms, charging, nursing rooms, bike parking, and more

Official reports

Formal complaints and suggestions sent to the TransJakarta CS dashboard
Automatic classification into 218 complaint subcategories
Numbered follow-up per complaint type: Tara only asks for what is still missing; users can answer all at once or step by step
Confirmation before submit, with the option to correct data

Track a report

Check report status with a ticket number (e.g. TJ-TIC05260004547)

Natural conversation

Greetings, thank-yous, and ending the chat naturally
Handoff to a human agent when the user asks to speak with CS
Polite handling of questions outside TransJakarta topics

Try It

Channel	How to access
Instagram	DM @infotije
WhatsApp	Customer Care Tara AI (official TransJakarta number)
TJ App	Customer Care Tara AI inside TJ:TransJakarta
Facebook	TransJakarta Messenger (smaller volume)
Web	Guest chat without login

My Role

Conversation & product logic: Built and extended the core chat flow for all intents above, including multi-turn conversations where Tara remembers what the user already said, across every channel.
Complaint reporting: Designed how Tara classifies complaints, asks the right follow-up questions per category, and sends a complete report to the internal dashboard, linked back to the original chat so agents can trace the conversation.
AI & knowledge base: Connected Tara to a searchable FAQ system so answers stay grounded in approved company content, and improved complaint classification so similar-sounding issues land in the right category.
CS team collaboration: Built tooling to auto-generate review spreadsheets from the official complaint taxonomy so CS staff can validate follow-up questions across 200+ complaint types without maintaining them by hand.
Admin dashboard: Worked on the internal web dashboard for session monitoring, sentiment overview, chat history review, and FAQ content management.

Technical Highlights

Complaint intake that mirrors how CS agents work

Before this work, Tara could guess the complaint type but only asked one round of follow-up. Agents still had to chase missing details like chronology, card number, or bus code. I separated "what type of complaint is this?" from "what do we need to ask for this type?" so CS teams own the question list while the product team owns the category structure.

When information is still missing, Tara sends a numbered list (e.g. 1. Chronology, 2. Card number, 3. Route). Passengers can answer everything at once or step by step. Tara skips questions already answered in the opening message. Location details (route, bus code, stop) are only asked when relevant. App-related complaints do not trigger bus-location questions.

Smart route help with two sources of truth

Passengers ask for directions in many ways: via Google Maps, "cheapest route," or simply "how do I get to Blok M?" I built logic that detects what the user wants and pulls from real-time map data when appropriate, or from an in-house route engine optimized for TransJakarta corridors when maps are not needed or not available. Responses clearly state which source was used so users are not confused when answers differ.

Impact & Results

The headline: fewer sessions, better outcomes

Overall volume went down, but quality went up. In June 2026, Tara handled about 21% fewer sessions than May. Yet deflection rate (conversations resolved without a human agent) rose from 82.8% to 89.6%. Tara needed human escalation less often, even with fewer total conversations.

By channel (May to June)

Channel	Volume change	Escalation rate
WhatsApp	Largest channel	28.3% to 17.1%. Biggest improvement and main driver of overall deflection gain
Instagram	Down ~29%	12.2% to 8.3%
TJ App	Down ~13% (stable)	6.2% to 3.5%. Best-performing channel overall
Facebook	Small volume	12.6% to 6.1%
Guest (web)	Up ~5x	5.1% to 1.6%. Low escalation, worth monitoring the traffic spike

Why escalation rates dropped

The most significant driver was the introduction of the AI-handled Report feature. Previously, any complaint submission required routing to a human agent — a reliable source of escalations across all channels. With Tara now able to handle reports end-to-end, a large category of sessions that would have automatically triggered escalation are now resolved by AI. This is most visible on WhatsApp, Tara's highest-volume channel, where escalation rate nearly halved from 28.3% to 17.1%.

Other results

218 complaint subcategories aligned with the official CS taxonomy
Numbered, multi-turn follow-up replaces a single generic question
CS review workflow with auto-generated spreadsheets for sign-off across all complaint types
Traceable reports where each submission links back to the original chat session
Deployed live across 5 channels, serving tens of thousands of sessions per month

Stack & Architecture

Layer	Choice	Why
Backend API	Python + FastAPI	Real-time chat from mobile and social channels
AI model	DeepSeek V3	Strong Indonesian language understanding at reasonable cost
Caching	Redis	Conversation state consistent across server instances
Database	PostgreSQL	Transit data, chat sessions, and analytics
Knowledge base	Vector search	Tara answers from approved FAQs instead of guessing
Route planning	Google Maps + in-house algorithm	Real-time directions when needed; TJ-optimized routing as fallback
Admin UI	Next.js + React	Internal dashboard for sessions, sentiment, and FAQ management

Lessons Learned

Early on, follow-up rules lived inside the main conversation logic. That worked for a handful of complaint types but broke down at scale. CS could not review or update questions without going through engineering. Giving CS a structured, reviewable checklist per complaint type was the real fix.

The June numbers confirmed the tradeoff was worth it: volume dropped, but quality rose. WhatsApp was the hardest channel in May (28% escalation) and saw the steepest drop. That showed better follow-up logic and classification matter more than raw session count.

I also learned that polishing AI responses for a friendlier tone can accidentally break app features. Confirmation buttons (Yes/No) disappeared when messages were rewritten. Formal report steps now keep a fixed format so the mobile UI always shows the right options.

TARA — TransJakarta AI Customer Service Chatbot