Mohit
Bansal.

Challenges

Coordinating a large multi-agent AI pipeline that had to support token-by-token streaming, RAG retrieval, and real-time collaboration simultaneously, without the different systems stepping on each other. Keeping generated content (reports, podcasts, debates, slides) reliably exportable and fully usable offline meant building a custom asset-caching and re-hydration layer. Real-time features (chat, presence, shared content, activity feeds) needed to stay in sync across a mobile app, an admin dashboard, and a public web app that all read from the same Supabase backend — while a credit-metered payment system had to fail fast and never silently lose a transaction.

Challenges

Balancing AI autonomy with safety in a mental-health context — the chatbot needed real-time crisis detection and a human-in-the-loop escalation path without becoming a bottleneck or a liability. Supporting 5 Indian languages meant more than translation: responses had to stay culturally and emotionally appropriate, not just linguistically correct. The forum needed anonymous posting robust enough that even the platform couldn't link identity to content, while still allowing AI moderation to catch harmful posts before they spread — all on a zero-PII data model built to satisfy GDPR and India's DPDP Act.

Challenges

Getting five different open-source models — a text model, a vision-language model, a speech recognizer, and a fusion layer — to reason together coherently on modest free-tier hardware, without the latency or memory footprint of a single giant multimodal model. Attention heatmaps and color-distribution analysis had to run in near real time to stay useful in an interactive demo rather than becoming a batch process.

Challenges

Keeping a rich-text document — tables, images, formatting, checklists — in sync across multiple simultaneous editors without corrupting structure or losing cursor context, while layering comments and mentions on top of content that could shift under a user mid-edit. Running on React 19 RC ahead of ecosystem support meant working around dependency incompatibilities across the whole stack.

Challenges

Merging two different market-data sources — Finnhub for live prices and Yahoo Finance for historical charts — into one consistent view without gaps or mismatched timestamps. AI-generated predictions needed honest confidence scoring rather than false certainty, and the portfolio simulator had to recalculate profit/loss in real time as live prices moved, without the UI feeling laggy under frequent updates.