About the Project
Inspiration
When deciding where to eat, grab coffee, or book a hotel, most people rely on star ratings and skim a few reviews. The problem is that star ratings flatten real experiences. A 4.5 score does not explain why a place is good, what trade-offs exist, or whether it matches what a specific person cares about.
The information people actually need is already in the reviews, but it is buried inside long, repetitive, and sometimes contradictory text. Reading dozens of reviews just to answer simple questions like Is it noisy? Is the service slow? Is it good value? is inefficient and frustrating.
I built Mieru to make reviews visible. Instead of summarizing everything into generic text, Mieru extracts what people consistently talk about and turns it into clear, evidence-backed insights that users can trust.
What I Learned
Building Mieru reinforced that review understanding is not just a language problem, but a system design problem.
I learned that:
- Pure LLM summarization is fluent but unreliable without structure and constraints.
- Trust comes from traceability, not from better wording.
- The most effective use of an LLM is not replacing the entire pipeline, but taking over the parts that require semantic judgment, while letting deterministic code handle stability and verification.
This project shifted my perspective from “prompting better” to designing guardrails around a model so that outputs are consistent, grounded, and usable for decision-making.
How I Built It
Mieru is designed as a simple but robust pipeline that separates concerns clearly.
1. Evidence Mining
Raw reviews are first cleaned and split into sentence-level evidence spans. Instead of directly asking the model to extract keywords, the system mines candidate phrases and experience-related signals using lightweight NLP techniques and TF-IDF ranking. This step narrows down high-signal content in a stable and reproducible way.
2. Gemini-based Consolidation
Gemini is used where it performs best: semantic understanding.
Given the top-ranked phrases and indexed evidence spans, Gemini:
- Consolidates semantically similar expressions into canonical concepts
- Assigns concepts to high-level experience aspects such as service, cleanliness, noise, or value
- Produces concise claims that summarize what reviewers consistently report
The model is constrained to output structured JSON and is instructed to rely only on the provided evidence.
3. Verification and Grounding
To ensure reliability, Mieru programmatically verifies that:
- Every claim includes at least one supporting evidence span
- Evidence references are valid and traceable
- Unsupported or overly generic claims are rejected
This transforms the output from a free-form summary into a decision-support artifact that users can inspect and trust.
Challenges I Faced
Balancing flexibility and stability
Handcrafted rules do not scale across domains, while unconstrained model outputs can be inconsistent. Finding the right balance between deterministic code and Gemini required multiple iterations.
Avoiding hallucinations without over-restricting the model
LLMs tend to generalize confidently. Designing prompts and verification logic that encourage honest uncertainty, rather than confident guesses, was one of the hardest challenges.
Designing for multiple place types
Restaurants, cafés, and hotels share similar review structures but differ in what matters most. The system needed to remain generic while still producing meaningful, domain-appropriate insights.
Why Mieru Is Different
Mieru does not aim to replace reviews or predict ratings. Its goal is simpler and more practical.
Make reviews visible by turning real experiences into grounded insights that people can quickly understand and trust.
By combining lightweight NLP, Gemini-powered semantic reasoning, and strict evidence verification, Mieru focuses on clarity, reliability, and decision usefulness rather than fluency alone.
Log in or sign up for Devpost to join the conversation.