Ayush Oswal

Product Engineer (Backend)

Azodha · Sep 2025 – Present · Pune, India (Hybrid)

Website

(They kept me. Interpret that how you will.)

• Cautiously migrated a critical microservice from PostgreSQL LISTEN/NOTIFY to a queue-based architecture, ensuring durability and zero downtime during the cutover — because nothing says 'career growth' like voluntarily touching a critical production service.
• Architected resilient n8n workflows to orchestrate custom client logic.
• Led the rollout of multilingual capability for the voice agent by evaluating and integrating multiple ASR/real-time providers (Deepgram, Gladia, OpenAI Realtime), reducing vendor lock-in and cutting new language onboarding time by ~60%, with support for real-time call transfers in case of clinical escalations. Turns out making machines understand humans in multiple languages is harder than it sounds. Surprising nobody.
• Integrated Customer.io for a client to power event-driven notifications, standardizing templates and tracking to improve notification delivery reliability and reduce time-to-ship new notification flows by ~50%.
• Implemented a resilient retry mechanism to handle silent queue disconnections, maintaining 100% availability across production workloads.
• Contributed to architectural decisions alongside senior engineers for onboarding new clients, translating custom business requirements into scalable, reusable service patterns.
• Designed randomized scheduling logic for batched notification delivery, distributing execution across time windows to eliminate concurrency spikes and prevent resource exhaustion under load.

SDE Intern (Backend)

Azodha · Dec 2024 – Aug 2025 · Pune, India (Hybrid)

Website

(Where the imposter syndrome began, and later, suspiciously, faded.)

• Assisted with the integration of OpenEMR and ECW systems into our platform for smooth data sharing.
• Developed custom logic for Welkin EMR, allowing automatic creation of 100ms links for patient sessions.
• Contributed to building Voice Agents with LiveKit, Deepgram, and OpenAI's realtime model for automating reminders and verifying users.
• Helped design and deploy RBAC middleware to protect patient information, maintaining HIPAA standards and blocking unauthorized entry in EMR setups. Protecting patient data: both ethically important and the kind of thing you want working correctly before anyone notices.
• Set up a token caching system for Lambda using DynamoDB to optimize expenses.
• Worked together on integrating Zoho Desk and CRM with our system.
• Created automated processes for handling Zoho ticket generation and scheduling recurring appointments.
• Engaged in production support and troubleshooting with DataDog and Signoz logs, aiding in fixing key problems.
• Improved database efficiency by examining Prisma queries via explain analyze, boosting response times by 20% through better indexing.
• Refactored legacy code to meet industry best practices and teamed up with experienced engineers on designing features and architecture.

(Self-appointed, naturally. The equity situation is very favorable.)

• Designed a serverless microservices setup with 3 AWS Lambda services via SAM templates and ECR, cutting infrastructure costs by 600% — because "we're running out of money" is a surprisingly effective architectural constraint.
• Created custom shell scripts to adapt Prisma binaries for Lambda compatibility and implemented AWS Lambda web adapter for streaming responses.
• Built an advanced RAG pipeline using JinaAI embeddings and Pinecone vector database for semantic searches, handling PDF/text with a 0.20 similarity threshold.
• Added a multi-LLM fallback system with DeepSeek as main and Gemini 2.0 as secondary to maintain service during rate limits.
• Developed a full token economy system providing 25,000 default tokens per user, tracking usage, and sending alerts through SMTP email.
• Implemented an API key-based chatbot integration with an analytics dashboard for monitoring interactions and easy website embedding.
• Incorporated SQS queues featuring retry mechanisms and DLQs for reliable, fault-tolerant operations.
• Introduced a contextualized chunking approach with 50 chunk overlap to preserve chunk relationships and enhance retrieval speed by 25%.