Santa Clara, California
I am a Software Engineer specializing in Agentic Systems & LLM Platforms at Adobe (Santa Clara, CA) on the 0→1 core engineering team building Adobe Brand Concierge (and previously Adobe Journey Optimizer). Joining roughly a month after initial product ideation, I have driven many of the foundational PoCs and technical spikes from prototype into production—spanning multi-tenant agent orchestration (PydanticAI, Model Context Protocol, A2A), grounded RAG with sub-second CPU neural cross-encoder rerankers, automated prompt and program optimization (GEPA, DSPy), and high-throughput streaming serving infrastructure.
Previously, I earned my M.S. in Computer Science at UC Irvine (UCI NLP Lab with Dr. Sameer Singh), researching in-context learning, prompt robustness, and chain-of-thought reasoning in LLMs (published at NAACL 2024 and EMNLP 2022), following a B.Tech at IIT Kharagpur. Earlier, I built distributed telemetry microservices at TruckX, search and analytics systems at Adobe (2018–2021), and cloud performance models at Adobe Research, co-authoring US Patent US20190354388.
Adobe Brand Concierge (BCOS) — 0→1 Multi-Tenant Agent Platform & Retrieval Infrastructure
Adobe Inc. • 2023–Present
PydanticAI • Model Context Protocol • ONNX Neural Retrieval • GEPA / DSPy
Core 0→1 engineering from early PoCs and spikes through enterprise production pilots (NVIDIA, Synopsys, ITC, DSG).
- Unified Multi-Tenant Agent Runtime: Replaced multi-hop legacy pipelines with a single PydanticAI agent runtime (BCOS) supporting declarative manifest compilation, A2A JSON-RPC, deterministic tool error recovery (
ToolErrorPolicy), SSE token streaming, and tenant/sandbox traffic isolation.
- Sub-Second CPU Neural Retrieval: Eliminated Candle batching limits by building CPU-only ONNX Runtime Text Embeddings Inference (TEI) services with dynamic axes—reranking 30+ candidate passages in a single batch at 588ms–796ms p50 latency alongside cross-encoder citation attribution (
bge-reranker-v2-m3, e5-curriculum).
- Automated Agent Optimization & Caching: Integrated GEPA and DSPy into Eval Studio for Pareto-optimal prompt compilation with 35x fewer rollouts than RL baselines, built a local FastEmbed semantic Q&A routing cache for sub-millisecond lookups, and filed three Adobe invention disclosures.
NAACL 2024
Selected
EchoPrompt: Instructing the Model to Rephrase Queries for Improved In-Context Learning
Rajasekhar Reddy Mekala, Yasaman Razeghi, and Sameer Singh
Findings of the Association for Computational Linguistics: NAACL 2024
PDF
arXiv:2309.10687
Abstract & BibTeX
We introduce EchoPrompt, a simple yet effective approach that prompts the language model to rephrase its queries before answering them. Just like humans, reminding the model of the question at hand by rephrasing it seems to help preserve context and reinforce connections across tokens. Through extensive experimentation with four causal model families across numerical reasoning (GSM8K, SVAMP), reading comprehension (DROP), and logical reasoning tasks, EchoPrompt yields substantial improvements (e.g., +5% on numerical reasoning and +13% on reading comprehension in Zero-Shot-CoT with code-davinci-002).
@inproceedings{mekala2024echoprompt,
title={EchoPrompt: Instructing the Model to Rephrase Queries for Improved In-Context Learning},
author={Mekala, Rajasekhar Reddy and Razeghi, Yasaman and Singh, Sameer},
booktitle={Findings of the Association for Computational Linguistics: NAACL 2024},
year={2024},
url={https://arxiv.org/abs/2309.10687}
}
EMNLP 2022
Selected
Snoopy: An Online Interface for Exploring the Effect of Pretraining Term Frequencies on Few-Shot LM Performance
Yasaman Razeghi*, Raja Sekhar Reddy Mekala*, Robert L. Logan IV, Matt Gardner, and Sameer Singh
Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: System Demonstrations (EMNLP 2022)
PDF
ACL Anthology
Abstract
SNOOPY is an online interface for exploring Pile term frequencies alongside precomputed model performance across NLP benchmarks. It lets users inspect how term frequency relates to few-shot accuracy across models, example counts, and datasets.
US Patent
Selected
US20190354388A1: Tenant-side detection, classification, and mitigation of noisy-neighbor-induced performance degradation
Subrata Mitra, Sopan Khosla, Sanket Vaibhav Mehta, Rajasekhar Reddy Mekala, and Aashaka Dhaval Shah
US Patent Application Publication US20190354388A1 (Adobe Inc., 2019)
Adobe Brand Concierge Invention Disclosures
- Demand-paged agent context with revocable tool capabilities
- Evidence-linked causal evolution of typed agent manifests
- Anti-collusion dual optimization for judges and agents
Rajasekhar Reddy Mekala (Adobe Inc.)
JOSS 2020
SunPy: A Python package for Solar Physics
Stuart J. Mumford et al., including Rajasekhar Reddy Mekala
Journal of Open Source Software, 5(46), 1832 (2020)