Pragya Mugale

Developed a privacy-first, full-stack meeting assistant for real-time transcription and intelligence, addressing data-sharing concerns of public commercial tools.

• Real-time STT Pipeline: Designed a high-concurrency WebSocket server in Go to handle duplex communication between frontend audio streams and Sarvam AI's STT engine.
• Microservices Orchestration: Auth, STT, and Meeting services containerized with Docker and ready for Kubernetes deployment, ensuring service-level isolation.
• Audio Processing: Engineered efficient binary chunking and base64 encoding strategies for high-fidelity audio streaming over WebSockets.
• LLM Integration: Built an asynchronous summarization engine using Gemini AI to generate structured insights post-meeting.

• Fault-Tolerant Streaming: Implemented session state management to handle abrupt WebSocket disconnections and resume transcriptions without data loss.
• Concurrency Management: Leveraged Go routines and mutexes for thread-safe audio buffering and multi-provider STT dispatching.
• OAuth & Security: Integrated robust authentication using Goth and Gorilla sessions, ensuring secure meeting access and data isolation.

• Treated the project with production-grade seriousness, implementing root cause analysis (RCA) for all bugs and maintaining a detailed bug log.
• Prioritized a clean separation of concerns between transcription, summarization, and core session management.

Designed a reliable activity logging system to capture high-volume user and system events across multiple products. Needed to handle write-heavy workloads with low latency while supporting concurrent high-frequency analytics reads.

• Write Path: Decoupled ingestion via a queue-based pipeline (RabbitMQ) to absorb traffic spikes and ensure 99.9% durability.
• Storage Strategy: Multi-database approach with DynamoDB for primary event persistence and ScyllaDB for optimized sequential reads/analytics.
• Read Path: Implemented independent read paths via internal gRPC/REST APIs, ensuring heavy batch processing doesn't degrade ingestion performance.
• Observability: Built exhaustive Grafana dashboards to monitor queue depth, consumer lag, and I/O latency.

• Asynchronous Ingestion: Prioritized availability and throughput by moving persistence logic to background consumers.
• Access Pattern Schema: Designed DynamoDB GSI/LSI strictly around application access patterns to prevent hot partitions.
• Idempotency: Guaranteed data integrity through unique event sequencing and deduplication at the consumer layer.

• Successfully handled write-heavy event ingestion with sub-10ms predictable latency.
• Improved system resilience through backpressure management and decoupled service layers.
• Enabled seamless downstream analytics without any production regressions.

Interested in working together or have questions? Feel free to reach out!