SAI CHILUKOTI
St. Joseph,MN
Overview
7
7
years of professional experience
Work History
Applied Scientist
uCTRL
01.2026 - Current
- Ported Google's ReAct-based Personal Health Insights Agent to AWS Bedrock with native tool-use and prompt caching; scaled 3,943 wearable-data questions across a Claude model ladder - Haiku 4.5 (90.6%) → Sonnet 4.5 (93.4%) → Opus 4.7 (94.3%) numeric accuracy - at 98%+ prompt-cache hit rate.
- Built and benchmarked a production RAG pipeline over the NutriChat human-nutrition PDF, evaluating fixed, semantic, and structural chunking with all-mpnet-base-v2 + Gemma-2B-IT; structural chunking outperformed baselines by up to 15% across RAGAS retrieval and generation metrics.
- Designed a defense-in-depth medical-LLM safety framework on AWS Bedrock - Guardrails input filter → frontier model → Claude Haiku 4.5 LLM-as-judge across 1,200 harmful prompts, denied-topic input filters added +36.7 pp block rate on GCG-jailbroken medical prompts (52% → 88.7%), and Sonnet 4.6 / Nova Pro eliminated worst-case direct-harm (0 score-5s) without medical fine-tuning.
Applied Science Intern
Amazon
Bellevue, WA
08.2025 - 12.2025
- Built a multi-modal SOP-generation pipeline for the Cybernaut web agent, unifying Chrome-recorder JSON, extracted video frames, and AWS Transcribe audio into Claude Sonnet 3.7-orchestrated prompts across four SOP variants benchmarked on an internal 25-domain suite + WebVoyager, the multi-modal JAV variant recovered task success on 13 tasks where JSON-only SOPs failed.
- Developed task-aware key-frame retrieval by fine-tuning a frozen DINOv2 backbone with a modified MoCo-v2 head using a dual queue and same-task hard negatives and improved the cross-task discrimination.
- Built semantic retrieval and wrong-action verification using Qwen3 embeddings, reranking, and similarity thresholds, selecting a 0.6B model with 12.07s embedding time versus 66.94s and 459.47s for larger variants.
Dissertation Fellow
University of Louisiana at Lafayette
Lafayette, LA
08.2024 - 08.2025
- Developed SMoE, a soft-gated Mixture-of-Experts model for multimodal glaucoma detection that mitigates modality imbalance via adaptive expert routing and load balancing, improving AUC by 1-3% over unimodal and balanced multimodal baselines across FairVision, FairDomain, and HarvardGF.
- Benchmarked DenseNet, ResNet, ViT-B, and EfficientNet on FairVision, then domain-adapted RETFound with MAE pretraining + LoRA fine-tuning, improving AUC from 81.16% to 83.2% and ES-AUC from 0.766/0.80/0.76 to 0.80/0.83/0.80 across demographic groups.
- Contributed to IdentityKD by collecting mmWave gait/facial data, implementing gait and KD baselines, and benchmarking against GaitSet, GaitPart, GaitGL, CRF, vanilla KD, CFKD, and MKD, achieving 98.58% accuracy and a 6.5% improvement over baseline.
- Contributed to MalInstructCoder by implementing baseline attacks, preparing poisoned instruction-tuning data, and benchmarking CodeLlama, DeepSeek-Coder, and StarCoder2, achieving 75-86% ASR@1 with only 0.5-1% poisoned data.
Research Assistant
University of Louisiana at Lafayette
Lafayette, LA
08.2023 - 08.2024
- Developed a computationally efficient transfer-learning and ensemble framework for 5-class diabetic retinopathy grading, achieving QWK scores of 0.901, 0.967, and 0.944 on EyePACS, APTOS, and Messidor-2.
- Developed DP-SGD-Global-Adapt-V2-S, combining adaptive noise and clipping schedules to improve privacy, utility, and fairness, achieving up to 4.01% accuracy improvement while reducing the privacy-cost gap by 89.83%.
- Part-time
Research Assistant
CVDI Research Institute
Lafayette, LA
08.2022 - 08.2023
- Developed a meta-learning-based federated learning framework with Global Prototype-Assisted Learning (GPAL) to reduce non-IID client drift and adapt to unseen tasks within 3 communication rounds, outperforming FedAvg and finetuning baselines on mini-ImageNet and tiered-ImageNet.
- Part-time
Research Assistant
Meta
08.2021 - 08.2022
- Developed privacy-preserving federated learning for fNIRS classification using DP-SGD, DP-Adam, and local differential privacy, achieving up to 98.2% accuracy while evaluating privacy-utility trade-offs across clipping, noise, batch size, and privacy budgets.
- Remote
- Part-time
Research Assistant
Velagapudi Ramakrishna Siddhartha Engineering College
Vijayawada, AP
01.2020 - 07.2021
- Developed a 41-layer modified ResNet in PyTorch for MSI/MSS classification on 192K histopathology images, achieving 89.81% accuracy and 91.78% F1, outperforming VGG16 and ResNet baselines.
Timeline
Applied Scientist
uCTRL
01.2026 - Current
Applied Science Intern
Amazon
08.2025 - 12.2025
Dissertation Fellow
University of Louisiana at Lafayette
08.2024 - 08.2025
Research Assistant
University of Louisiana at Lafayette
08.2023 - 08.2024
Research Assistant
CVDI Research Institute
08.2022 - 08.2023
Research Assistant
Meta
08.2021 - 08.2022
Research Assistant
Velagapudi Ramakrishna Siddhartha Engineering College
01.2020 - 07.2021