Akshat Barve

01The problem

Health reports pile up as disconnected PDFs. One from this lab, another from that hospital, each using different names for the same test. Patients can't track trends across reports without manually cross-referencing values. Most health apps ask you to type in your own data or don't handle the inconsistency between lab naming conventions at all.

02The approach

AarogyaKul runs a six-stage async pipeline behind a single 202 Accepted response. Upload a blood report PDF and the system tries direct text extraction first, falling back to Tesseract OCR only when needed (10-50x faster than OCR-everything). A Llama-based LLM extracts structured values, a 26-entry canonicalization map normalizes lab naming variations, and a BigDecimal-based trend algorithm compares against prior reports with zero floating-point drift on medical values.

01The problem

AWS cost spikes don't announce themselves. A misconfigured resource or a runaway process can compound for weeks before anyone notices on the monthly bill. By the time the overrun surfaces, the damage is already done. Most teams rely on billing alerts with static thresholds that either cry wolf constantly or miss the subtle spikes that matter.

02The response

Five single-responsibility Lambdas, decoupled via EventBridge, form a detection-to-alert pipeline. A two-table DynamoDB schema (raw vs. pre-aggregated, with a GSI for trend queries) feeds a Z-score + spike-based detection layer. Thresholds are cold-start-aware so the system doesn't cry wolf during a new account's first few days. Anomalies route straight to Slack with a deep link, from signal to a human's phone before the spike compounds further.

01The problem

Keyword-based resume screening misses qualified candidates whose experience is described differently. A 'Backend Engineer, JVM ecosystem' resume gets rejected when the search is 'Java Developer'; same domain, zero match. Recruiters either review everything manually or accept that automation means losing good candidates to phrasing mismatches.

02The shift

A Spring Boot 3.0 REST API replaces keyword matching with Meta Llama-3 semantic analysis. The system understands that 'JVM ecosystem' and 'Java Developer' describe the same domain. Async parsing decouples AI inference from the request lifecycle; the API returns an immediate 200 OK while Llama-3 processes in the background. Redis caching and optimized PostgreSQL indexing keep API latency under 200ms with 90% match accuracy.