One Month into GSoC with AOSSIE: My Journey Building Perspective

#gsoc #opensource #webdev #langgraph

🚀 TL;DR

I’m one month into Google Summer of Code with AOSSIE, building Perspective- Perspective analyzes your news or social feed and presents credible counter-narratives from reliable sources-helping you think critically, reduce bias, and see the full picture. Don’t settle for one-sided stories. Get complete, nuanced facts.

🎯 1. Project Overview

Perspective analyzes news articles to surface hidden biases by combining cleaning, entity extraction, and bias scoring. I’m collaborating with mentors Manav, Pranavi and Bruno on the Perspective repo.

📆 2. Month 1 Milestones

Week	Deliverable	Status
1	Project setup & Figma mockups	✅ Done
2	Frontend pages (Landing, Loading, Results)	✅ Merged PRs
3	`cleaner.py` & `extractor.py` modules	✅ Merged PRs
4	UV‑based backend, scraper module, LangGraph	✅ Merged PRs

📔 3. Day‑by‑Day Highlights

June 2–5:

– Revamped frontend: Landing & Analyze pages in Next.js+Tailwind.

– Built Loading & Results screens with global state flow.

– UI/UX tuning and responsiveness fixes after mentor feedback.
June 6–10:

– Squashed a React‑date‑picker build bug.

– Finalized frontend deployment on Vercel.

– Explored uv for Python dependency management; set up new backend architecture.

– Added a FastAPI scraper module for article ingestion.
June 11–15:

– Dug into LangGraph: designed the pipeline flow (Sentiment → Fact‑Check → Generate → Judge → Retry).

– Implemented base StateGraph and stubbed out node files (sentiment.py, fact_check.py, etc.).

– Researched error‑handling patterns and added an ErrorCatcher node.
June 16–20:

– Integrated Groq’s Python SDK for sentiment analysis; tuned prompts for deterministic outputs.

– Built DuckDuckGo search node and LLM‑analysis node for real‑time fact‑checking.

– Crafted structured CoT prompts and wired up the generate_perspective node.
June 21–25:

– Developed & integrated judge_perspective scoring node (originality, reasoning, factual grounding).

– End‑to‑end tests on real articles (e.g., 2025 French Open final).

– Refactored for stability: added fallbacks, cleaned up exception flows.
June 26–30:

– Chunking utility for vector DB ingestion; set up Pinecone embeddings & metadata storage.

– Linked frontend to backend for fact‑check results; switched from Render to Hugging Face Spaces for deployment.

– Automated backend CI/CD with GitHub Actions; local testing via act.
July 1–7:

– Polished docs & example pipeline states; pushed final PRs.

– Monitored deployments, fixed minor bugs, and synced progress with mentors.

🛠 3. Deep Dive: What I Built

3.1 LangGraph Workflow

Challenge: While building LangGraph, efficiently handle all the data and making sure to pass right state takes lot of efforts.
Key Learnings: LangGraph workflow needs logging state after each node and robust error handling.

🤯 4. Challenges & Learnings

LangGraph quirks:
- I hit a bug when chaining async nodes. Fixed by adding explicit await and restructuring my graph definition.
- Tip: Write unit tests for each node in isolation—saved hours of debugging.
uv in deployment :
- uv needs carefull docker setup for deployment.
- Fix: loaded uv in container and ran my main.py with --no-cache paramenter to prevent container permission issues.
Collaboration flow:
- Coordinating with three mentors across different time zones taught me to write crystal‑clear PR descriptions & use GitHub Projects for tracking.

🚀 5. What’s Next (Month 2)

Bias Scoring Module: Prototype a heuristic‑based scorer combining sentiment analysis + entity context.
Vector DB Integration: Push cleaned articles & metadata into Pinecone for fast similarity queries.

I’ll be posting updates every 2–3 weeks-stay tuned!