Making A Peer Review System for My Blogs Using Google-ADK & Mem0

My Process

When writing my technical blogs, I have a very rigid process I like to follow.

1. Research the topic I am interested in

2. Create a structured research roadmap I require to gain knowledge about the particular topic

3. Go through the roadmap and try to learn/research the concepts as in-depth as I can

4. Start coding whatever the relevant implementation for that topic is

5. Finally, start writing the blog

But one thing always bugs me, "Is my blog factually correct and have I compromised the integrity of my blog anywhere?".

That leads me to frantically go through my sources repeatedly and asking tools like Perplexity about the blog. So, I had the idea to automate this process by a creating a Peer Review System.

---

What This System Does

1. What the System Focuses On

• It behaves like a technical editor, not just a grammar checker.

• It evaluates writing for:

  • Structure

  • Clarity

  • Factual accuracy

  • Tone correctness

  • Proper use of supporting evidence

• The purpose is to help the writer produce content that is accurate, readable, and consistent.

2. How It Reviews Content

• The system doesn’t read content blindly.

• It uses uploaded reference files as a knowledge base.

• Relevant information from those files is retrieved using semantic search rather than keyword matching.

• If a statement appears in the writing:

  • The system first checks if it exists in the uploaded sources.

  • If confirmed, the system becomes more confident in that claim.

  • If not found, it triggers an external web-based fact check.

3. How Memory Improves Review Quality

• Feedback adapts over time instead of resetting with each review.

• The system tracks repeated mistakes or patterns such as:

  • Missing citations

  • Style inconsistencies

  • Formatting issues

• If the same issue shows up again, the system highlights it more firmly.

• This turns the review into a learning process rather than a one-time correction.

---

Screenshots:

• https://drive.google.com/file/d/1VTvQBkQ4753NbpVFlVPr5v6_kjH3SXjZ/view?usp=sharing

• https://drive.google.com/file/d/14sfHrC0Lw0pvU4oX7U18Ydv61a8RNTu0/view?usp=sharing

A Demo Peer Review Report:

• https://drive.google.com/file/d/1HQh5stEAj4tkh3E7Fyf1jOse52ZDw-bb/view?usp=sharing

---

Workflow

Phase 1: Ingestion

• Fetches content from URLs if needed

• Loads past review history for the project

• Examines uploaded source documents

Phase 2: Verification

• Identifies all factual claims in the content

• Searches uploaded sources for supporting evidence

• Uses Google search for external fact-checking

• Validates technical assertions and statistics

Phase 3: Evaluation

• Assesses clarity, flow, and structure

• Checks accuracy against evidence

• Evaluates tone for target audience

• Compares to past feedback to track improvement

• Flags recurring issues with escalated severity

Phase 4: Synthesis

• Generates structured report

• Provides evidence for all major issues

• References past feedback when relevant

• Gives actionable, constructive feedback

---

Features

1. Model Flexibility

• You aren’t locked into one AI provider.

• Switching between models like Gemini, Claude, GPT, or Ollama only requires changing one environment variable.

• This gives control over:

  • Cost

  • Performance

  • Privacy

• The review logic remains consistent across models.

2. Context-Aware Retrieval

• Uploaded reference files are stored in a vector database.

• The system breaks them into chunks, embeds them, and indexes them for efficient search.

• During review, it retrieves relevant sections using semantic similarity rather than simple keyword matching.

• This helps the system understand meaning, not just matching exact text.

3. Automated Fact Verification

• When a claim isn’t supported by uploaded sources, the system escalates verification.

• A separate search agent performs a structured web lookup.

• The goal is not to rewrite content, but to confirm whether the information is reliable and accurate.

4. Built-In Memory

• The system remembers past reviews and writing patterns.

• If a mistake repeats, the system identifies it as a recurring issue.

• Instead of pointing it out repeatedly at the same level, the feedback becomes stronger and more specific.

• This encourages long-term improvement rather than one-off corrections.

---

Limitations

1. The verification is only as good as the model plus the search results

2. Source reliability isn’t enforced

3. Web search can surface low quality or outdated material

4. The model is still the final judge. It can misinterpret sources, over trust weak evidence, or fabricate justification

Implementation: GitHub