1 of 23

Lotara

2 of 23

Agenda

01

02

03

04

05

06

February, 2026

Team members

Pain points

Solutions

Challenges

Demo

Next plan

3 of 23

Team members

Khoi Nguyen

Leader, Backend Developer

Man Duong

Frontend Developer

Hung Tran

AI Developer

Phuc Tran

Data Analysis

February, 2026

4 of 23

5 of 23

One-Size-Fits-All Itineraries

Tourists take a boat tour while visiting Hoi An City in Quang Nam Province, central Vietnam. Photo: Nam Tran / Tuoi Tre

February, 2026

Tour packages treat every traveler the same, with fixed schedules and generic activities that ignore personal interests and pace. Travelers end up rushing through places they love and spending time on experiences they don’t care about.

6 of 23

Solution

Lotara is a personality-driven travel platform that uncovers hidden local gems through data and helps travelers plan affordable, meaningful trips.

7 of 23

8 of 23

9 of 23

Architecture

10 of 23

11 of 23

12 of 23

Comet Opik

13 of 23

14 of 23

15 of 23

Next plan

Optimize waiting time

More features

Enhance search

16 of 23

Thank

you

17 of 23

Pain points

Overcrowded, Repetitive Destinations

Hidden Gems Are Hard to Find

One-Size-Fits-All Itineraries

February, 2026

Travel companies promote the same famous places, leading to crowded and predictable trips.

Authentic local spots are scattered across the internet, outdated, or difficult to access.

Tour packages ignore personal interests, pace, and travel style.

18 of 23

Challenge

Multiagent architecture

RAG

Request timeout

Outline what needs to be done to prevent orchestration loops and conflicts.

Irrelevant document chunks occasionally overwhelmed the context window, leading to decreased response precision.

The cumulative time of sequential LLM reasoning steps frequently exceeded the standard roughly 150-second gateway threshold.

Difficulty in tracing errors across non-linear workflows when multiple agents operate concurrently.

Managing handoffs between specialized agents led to redundant tool calls and inconsistent state management.

Standard embedding models struggled to capture the specific domain nuances required for highly accurate grounding.

Challenges in maintaining a real-time vector index as the underlying source data was updated.

System performance degraded under high load, causing API timeouts during peak multi-agent processing.

Heavy prompt engineering and complex system instructions increased token processing time beyond acceptable UX limits.

19 of 23

Technology

Frontend: NextJS

Backend: NestJS, FastAPI

Agent Development Kit: Google ADK

Database: Postgres

Vector Database: Milvus

Model: Gemini

20 of 23

21 of 23

Google Gemini

1

2

3

4

Gemini 2.5 acts as the core "brain" for our multi-agent system, providing the reasoning depth and context window (1M+ tokens) required to plan and execute complex, multi-step workflows. It transitions the system from simple responses to autonomous task completion.

Deploy Gemini 2.5 Flash as the master orchestrator to manage state across 1M+ tokens, ensuring no information is "lost" during long agent conversations

Offload sub-tasks (like travel inspiration, planning or repetitive tool calls) to Gemini 2.5 Flash to achieve sub-second latency and reduce operational costs.

Tasked with managing API rate limits (up to 2,000 RPM for Flash) and ensuring high availability of the multi-agent system.

Gemini 2.5 serves as the central engine within the ADK framework, allowing specialized agents to learn and collaborate in real-time to synthesize a single, cohesive response.

22 of 23

Comet Opik

00%

00%

Key performance indicator

Key performance indicator

To implement a robust observability layer that moves beyond "vibe checks" to data-driven agent optimization. We will use Comet Opik to trace multi-agent reasoning, monitor token costs, and manage prompt iterations in a unified library.

Use Opik’s Trace and Span structure to reconstruct the full decision path. A Trace represents the entire user request, while Spans break down individual tool calls, retrieval steps, and agent-to-agent handoff.

Monitor total token usage broken down by phase (planning vs. tool-use vs. final response), source (system prompt vs. RAG context), and model tier (Gemini model)

Observe the failure rate (null or error responses) and set up alerts for agent loops—where agents repeat the same tool call with identical parameters

Directly visualize agent trajectories—showing exactly how an agent planned, which tools it chose, and the reasoning behind each decision.

23 of 23

Timeline

Week 1:Getting Started

Week 2: Halfway There!

Week 3: Keep Building

Week 4: Submission Time

Build team

Brainstorming ideas

Collect and clean destinations

Core features implemented: User onboarding, Recommend places (without AI)

RAG with vector DB

Prepare demo / documentation