Aman Kaushik

● Engineered a production-scale knowledge graph and LLM pipeline (Neo4j) to automate OCR-based content structuring. Solved complex data extraction bottlenecks for major educational curricula (NCERT, Maharashtra and Odisha State Boards, Bansal Classes), standardizing unstructured enterprise data at scale. ● Built production-grade pipelines for document classification and multilingual translation for the Maharashtra Government (Tapaal). The system processes thousands of daily inquiries across regional vernaculars, automating document routing and drastically reducing manual triage time for state-wide operations. ● Orchestrated agentic multi-agent workflows for EXL (Fortune 500), automating intricate business logic through modular agent design. Leveraged advanced state management and agentic frameworks to deliver high-reliability, mission-critical automation for enterprise-scale operations. ● Architected a provider-agnostic multi-modal engine for automated video, audio, and image synthesis, powering a global commercial rollout on Sumo AI. Built a polyglot backend (Python/TypeScript) made for horizontal scalability, successfully navigating high-concurrency traffic surges from worldwide customers while maintaining strict feature parity across disparate APIs. ● Deployed observability and evaluation pipelines using LangSmith and Opik across 5+ production client projects. Tracked hallucination rates, classification precision, and translation quality to enable dataset creation and iterative prompt engineering. ● Implemented scalable Python backend services using FastAPI and Dapr on GCP, implementing async processing and caching for AI-heavy pipelines serving thousands of document processing requests across active client deployments.

● Fine-tuned YOLOv8 models on large-scale public datasets for document and face processing achieving 0.95 mAP in inference. ● Designed FastAPI services for liveness detection and document processing pipelines with 98% classification precision.

● Built a multi-agent AI system for global supply chain analysis using Neo4j GraphRAG , combined 4+ retrieval strategies (vector, hybrid, vector+cypher, hybrid+cypher) to extract insights from complex structured and unstructured data. ● Applied Voyage-3 embeddings, BM25, and custom re-ranking with Agno AI. Developed a fully asynchronous architecture with SSE streaming, achieving a TTFT of under 1.5 seconds.

● Orchestrated a multi-agent travel concierge using Agno (Phidata), coordinating 6 specialized agents (Amadeus, Perplexity, Google Maps, NewsAPI, TripAdvisor, ElevenLabs) under a Claude 4 Sonnet team lead. Integrated SSE streaming, tool-call tracing, and runtime context injection. ● Engineered an episodic memory service using LLM-based structured summarization and pgvector similarity search for cross-session personalization. Created an automated ElevenLabs audio-tour generator to produce interactive, playable travel guides.