CV

• Designed and deployed a production-grade multi-agent orchestration engine for workflow automation, enabling natural-language interaction with 4000+ API integrations and custom agents. [Patent Pending]
• Mitigated operational risk with a security-first execution layer featuring credential-scoped tool execution, human-in-the-loop (HITL) approval gates for high-stakes API actions, and full execution traceability.
• Engineered a hierarchical orchestration engine with long-term memory and RAG-based tool discovery, enabling the system to handle complex, multi-turn tasks while minimizing context bloat.
• Developed data augmentation pipelines for API documentation and tool metadata, generating provider-specific agent skills and improving tool retrieval quality.
• Designed an evaluation framework to benchmark LLM planning and tool-use capabilities across standard and edge-case scenarios, enabling regression testing and model drift monitoring over time.

STACK: AWS Cloud, Amazon Bedrock, OpenAI, Gemini, Mistral, LiteLLM, RAG, multi-agent orchestration.

• Investigated systematic compositional generalization in transformer-based seq2seq models by incorporating syntactic structure into the decoding process through hyperbolic representations of dependency trees.
• Designed, trained, and evaluated a hybrid Euclidean–Hyperbolic transformer architecture for structure-aware sequence generation.

STACK: Transformers, Fairseq, PyTorch, Geoopt.

• Studied advancements with the ColBERT architecture, which relies on token-level representations with late interactions for document ranking.
• Proposed a structured distillation approach for ColBERT contextualized token embeddings using aggregated frozen pre-trained term-topic embeddings representing token-level contextual semantics.

• Developed a tweet summarization pipeline leveraging deep learning models for semantic tweet representation and relevance ranking with respect to user interests.