Venu
Madhav

Built a system that lets multiple organizations search each other's data without exposing sensitive content, using a federated RAG pipeline with FAISS, BM25, and cross-encoder reranking.

Implemented 4 privacy layers — embedding noise (Differential Privacy), score encryption (Homomorphic Encryption / CKKS), binary hashing (LSH), and secure aggregation — preserving 94% retrieval accuracy while fully hiding individual scores. Tested defenses against 4 real-world attacks (membership inference, score leakage, embedding reconstruction, query fingerprinting); all score-based attacks dropped to random-guess accuracy.

Engineered a production-style Retrieval-Augmented Generation (RAG) system for syllabus-grounded question answering, supporting multimodal ingestion (PDF, PPT, video transcripts) and semantic retrieval using FAISS and Qdrant.

Built modular RAG pipelines with LlamaIndex and MistralAI, with optional Kafka-backed streaming. Implemented a fully local RAG evaluation framework (470+ queries) using TF-IDF and cosine similarity to measure context precision, recall, and answer correctness. Real-time chat architecture via WebSockets and Redis session management.

Built a Kafka-inspired distributed message broker implementing leader-based replication, WAL-backed durability, and offset-driven log consumption. Extended it with real-time transformer-based sentiment inference.

Redis lease-based leader election (5s TTL) with synchronous high-watermark replication. Consumer-managed offsets for at-least-once delivery. Bloom filter deduplication (O(1) lookup) and zero-downtime model switching via Redis.

Designed a control-plane + data-plane Kafka architecture enabling dynamic topic lifecycle management (pending → approved → active → inactive → deleted) with Admin-driven approvals, broker-side topic creation, and real-time producer/consumer coordination.

Built a multi-threaded producer (publisher, input listener, topic watcher) with Kafka Admin API integration. Implemented database-backed topic governance using MySQL for metadata and subscription control. Enabled dynamic consumer subscription to active topics with auto.create.topics disabled for strict control. Centralized configuration via JSON for broker, replication, offsets, retries, and poll tuning.

Designed and implemented a fully normalized (3NF) relational database system for gym management, exposing business logic through FastAPI. Engineered schema with strong referential integrity, transactional workflows, and production-style reporting queries.

Built 14 relational tables with constraints, indexes, and optimized JOIN paths. Implemented 5 triggers, 7 stored procedures, and 4 SQL functions for automation and business rules (membership activation, capacity control, coupon logic). Developed 20+ complex analytical queries covering revenue, retention, utilization, and activity reports.

Discord bot for server moderation with automated content filtering, user management, and administrative commands for maintaining community standards.

Automated moderation with keyword filtering. User warning and ban management. Custom command handlers, event listeners, and server analytics with logging features.

Developed a canvas-based AI math assistant that converts handwritten equations into structured mathematical expressions using OCR and generates step-by-step solutions through LLM inference.

Integrated Google Flash 2.0 API for equation recognition with custom canvas preprocessing. Built a real-time React + Node.js architecture for image capture, API orchestration, and dynamic solution rendering. Implemented structured prompt engineering to ensure coherent, step-wise mathematical reasoning.