Balachandra Devarangadi Sunil

Accomplishments

• Achieved up to 90.9% latency improvement (11x speedup) on individual ResNet50V2 partitions and drove the system-wide result of up to 63.8% mean latency reduction for single-tenant and 77.4% for multi-tenant workloads versus the default Edge TPU compiler.
• Diagnosed and solved a hard cross-stack correctness problem, inconsistent TPU/CPU execution results traced to TFLite versioning, restoring consistent output and unlocking a validated 24% latency improvement for InceptionV4 pipelines.
• Built a dual-threadpool inference pipeline (separate TPU and CPU thread pools with closed-loop request handoff) to enable accurate throughput benchmarking across split points and CPU thread counts, and generated model-size and partition tables that directly informed the paper's analytic queueing model design.

Tools

Accomplishments

• Built an ensemble architecture where a meta-LLM selects the best answer from structured Chain-of-Thought candidates generated by each retrieval method.
• Implemented dual retrieval backends (BM25 & Elasticsearch + FAISS) to support flexible sparse and dense retrieval across different invocation strategies.
• Solved the challenge of dynamically deciding which retrieval strategy fits a given query, unifying multiple RAG invocation methods into one adaptive, ensemble-driven system.

Tools

Accomplishments

• Designed 32+ conversation stories and NLU training data covering intents like NPA tracking, loan disbursement, regional sales performance, and live loan counts, each with multi-turn slot-filling dialogue flows.
• Integrated Duckling's time-entity extraction to resolve natural language expressions (e.g. 'last quarter', 'same month last year') into structured timestamps for backend queries.
• Built custom Rasa SDK actions that mapped extracted entities and slots (type, region, time, order) into structured backend API calls, with automatic slot-resets and fallback handling for robust conversation recovery.

Tools