AbhinavSaxena/buildsthings thatdon't break.

The fix was an event-driven microservice with a smart multi-channel cascade — email first, then WhatsApp, then SMS, then an AI-placed phone call — with circuit breakers on each. Distributed locks on Redis made sure no two workers picked the same candidate. Idempotency keys made retries safe. EXPLAIN-guided index work shaved 45% off average query latency.

Result: TAT collapsed from 5 days to 1 day, per-candidate cost fell from ₹500-800 to ₹200-300, and the team started saving ~₹6L every month.

We built an end-to-end AI pipeline: Llama Parse and Llama Extract for structured field extraction, Claude and GPT-4 for classification and auto-grading, Portkey-mediated semantic caching for cost control, and a Kubernetes-based worker tier that auto-scales with priority queues.

Consistency across services during the migration was preserved via the Saga pattern on Kafka. Auto-grading rate jumped from 0% to 30%; TAT dropped from 3-4 days to 1-2.

Users drop into topic-specific rooms — trauma, grief, anger management, social skills — and speak freely without identity attached. Volunteer listeners pair up for 1:1 audio. Hugging Face emotion classifiers run on spectrograms extracted client-side; the round trip stays under 2 seconds.

Privacy was the hardest constraint: audio is WebRTC-encrypted end-to-end, stored temporarily on encrypted S3 with 24-hour auto-deletion, and only emotion metadata is retained long-term.

Storing medical files on-chain is prohibitively expensive, so Vaastavik takes a hybrid approach: files live on IPFS; only their content hashes (≈46 bytes) are written to Ethereum. A multi-tier admin system validates hospitals via license documents before they can ever join the network.

The tricky part was making blockchain UX invisible. MetaMask prompts became clear, one-click confirmations; Web3 calls got abstracted behind a friendly React layer so medical staff never had to think about gas.

A Next.js frontend calls SerpAPI for Scholar results. A FastAPI backend scrapes selected papers with headless Chrome, extracts structured content with CSS selectors, and hands it to GPT-3.5-Turbo as context for grounded Q&A.

Bot detection was the ongoing cat-and-mouse: rotating user agents, careful error handling for JS-rendered content, Docker-packaged Chrome with remote debugging — all so students could ask "what is the methodology?" and get a useful answer in seconds.