Madhu Sudhan Rao Korivi

- [Go] Built a Pion WebRTC peer-to-backend media server — browser clients stream video frames to the Go server via SDP/ICE/DTLS signalled over REST APIs; server buffers received RTP frames and publishes them as Kafka events for downstream AI processing, keeping real-time communication fully decoupled from AI workloads. - [Go/Python] Architected Apache Kafka event backbone for media frames, and AI insights: Go producers with exactly-once semantics and message batching; Python consumers with idempotent writes, DLQ routing, and consumer group coordination — high-throughput event pipeline with sub-500ms Kafka processing latency. - [Python] Owned LangGraph agentic workflow engine: multi-step autonomous agents with tool calling, memory management, and conditional branching; provider-agnostic LLM routing across OpenAI GPT-4o, Gemini 2.5 Flash, Groq Llama3, and Anthropic Claude via Strategy Pattern abstraction. - [Python] Built RAG pipeline with ChromaDB and pgvector: document chunking, batch embedding (Gemini Embedding-001), semantic similarity search, and hybrid retrieval piped into LLM inference with exponential backoff retry. - [Python] Deployed Apache Flink jobs to ingest Kafka-processed events into PostgreSQL for analytics dashboards, decoupling analytical workloads from the live event pipeline. - [Go/Python] Designed and shipped an internal OpenTelemetry SDK (separate implementations for Go and Python) abstracting trace propagation (W3C TraceContext across HTTP and Kafka headers), custom span attributes, Prometheus metrics, and structured log correlation; integrated with SigNoz — MTTR reduced by 50%. - [Python] Developed FastAPI/Django REST APIs with Pydantic validation, JWT/OAuth 2.0 auth, and OWASP Top 10 controls; optimised PostgreSQL with advanced indexing and PgBouncer connection pooling — query latency reduced by 45%, write throughput up 30%.

- [Go/Python] Built and maintained REST APIs in Go and Python (FastAPI) for ESG platform services with structured logging, request tracing, and error handling baked into every layer; implemented AWS S3 integration for storing and retrieving ESG documents and images — pre-signed URL generation, multipart upload, and lifecycle policies for cost-optimised storage. - [AWS] Used AWS DynamoDB (Python Boto3 SDK) to store document metadata and audit records with efficient key design for fast lookups; contributed to EC2-based staging environment configuration for internal deployment. - [Python] Improved API response time by 62% (250ms → 95ms) via FastAPI async I/O, SQLAlchemy 2.0 async drivers, and Redis caching; redesigned MySQL/PostgreSQL indexes (EXPLAIN ANALYZE) across 20+ endpoints — 60% throughput gain. - [Python] Integrated LLM APIs (OpenAI, Gemini) with structured JSON output mode and prompt engineering for automated ESG report classification, entity extraction, and summarisation pipelines.

ChatPDF — Production RAG Document Assistant

September 2022

Fully async FastAPI RAG app: aiofiles non-blocking PDF I/O, thread-pool pdfplumber parsing, native batch embedding API, exponential backoff retry, provider-agnostic Factory (Gemini/OpenAI for LLM and Embeddings), ChromaDB cloud/local, Pydantic strict validation, JSON structured logging.