Qyudal

Automate the analyst workflow: 7 AI agents that ingest, explain, predict, fact-check & report on the news in under 4 seconds.

Comment

Inspiration

It started with the realisation of my endless scroll. One night, deep in an Instagram reels spiral, it hit me that I was just consuming brainrot for hours. So I switched platforms and downloaded a news app; Inshorts to feel a little less guilty. But the news didn't hook me either, and worse it I felt it an imposition; the viewpoints, they clearly appeared biased. One headline, one angle, one implied opinion.

That's the real problem. News isn't black or white ;it's a shade of gray. Which side you land on should be YOUR call as the reader, not something quietly imposed by whoever wrote the summary. But getting the full picture means manually hunting across sources, weighing conflicting viewpoints, and fact-checking as you go slow, tedious work that almost nobody actually does.

I built this app to automate exactly that workflow.

What it does

Qyudal is a multi-agent news intelligence engine. Most news apps hand you a single, often one-sided take. Qyudal does the analysis for you: for any story, seven AI agents run in parallel they explain it, predict where it's heading, debate it from opposing sides, build a timeline, visualize it, take notes, and fact-check it by cross-checking the story across multiple sources for corroboration and return the full picture in under four seconds. So instead of a pre-chewed, biased snippet, you get every angle, sourced and scored, and you decide where you stand. Each morning, Qyudal auto-compiles the day's top stories into a single brief the whole news-analysis workflow, automated. It's my submission for the Artificial Intelligence & Machine Learning track, with a strong Social Good dimension improving news literacy for 4.4 billion underserved Gen Z users globally.

How I built it

Built for the AI/ML track, NewZ demonstrates multi-agent orchestration, LLM inference optimization, and production-grade scalability. Ingestion- NewsAPI for live, multi-source story feeds across topics. Intelligence-Groq for fast LLM inference, powering 7 purpose-built agents (Explain, Predict, Debate, Timeline, Visualize, Notes, Fact-Check) orchestrated to run in parallel rather than sequentially. App- A React Native frontend for a fast, mobile-first experience. Auth & state- Firebase for secure login and persistent sessions, so analysis and saved stories carry across devices.

Challenges I ran into

Multi-agent latency- Running 7 agents sequentially caused 10-20s delays;unusable. therefore, parallelized the calls with Promise.all(), cutting total analysis time to under 4 seconds. API rate limits - Parallel category queries hit rate limits and returned duplicate/stale data. added an in-memory cache plus page-based pagination to spread the load. Prompt reliability - LLM responses returned inconsistent JSON, breaking the UI unpredictably. enforced strict JSON-schema prompting with robust fallback parsing. Mobile performance - Animated card transitions dropped frames on lower-end Android devices; fixed it with the native-driver Animated API and FlatList windowing for smooth 60fps. Content quality - Early summaries were too verbose and formal not decision-ready. hence, tuned prompts for a concise, scannable tone capped at two lines, with a confidence score on every summary.

Accomplishments that I proud of

Put it in front of real people that includes peers, professors, and UPSC aspirants(my relative) and every group said it genuinely addresses a problem they face. For aspirants especially, daily news analysis is hours of manual work, and they immediately saw the value. A 7-agent pipeline returning full analysis in under 4 seconds, from a standing start. A genuine multi-viewpoint Debate feature surfacing opposing perspectives instead of a single imposed take.

What I learned

Orchestration beats raw model power.The single biggest win wasn't a better prompt it was running agents in parallel. Architecture decided whether the product was usable. Reliability is a prompt-engineering problem.Schema enforcement and fallback parsing mattered as much as the "intelligence" itself. Validate early, validate with real users. Watching aspirants and researchers react told us more than any internal assumption and confirmed we were solving a real workflow, not an imagined one.

What's next

Multilingual support : Hindi, Tamil, and Bengali to reach beyond English-first readers. Audio briefings: AI voice digests for people on the move. Publisher API partnerships: direct, licensed feeds from trusted outlets. Nexus Intelligence: a cross-story entity-relationship graph that connects how events, people, and organizations link over time. Smart notifications: agents that decide what you need to know, not everything. for Enterprise: team workspaces, role-based briefs, and audit logging for newsrooms, research, and analyst teams.

ppt-

View Presentation

Built With

Share this project:

Updates