Harshith Sai Peram

• Built and operated large-scale simulations of end-to-end Amazon fulfillment flow, from inbound receiving to last-mile delivery, to forecast customer demand and warehouse staffing needs for the next 16 weeks, enabling scenario planning and capacity decisions across sites.
• Designed robust data pipelines to ingest and unify high-volume, multi-service datasets; implemented preprocessing (schema enforcement, deduplication, anomaly checks, late-arrival handling) to ensure trusted inputs for simulation and analytics.
• Created a reproducible simulation data layer with versioned inputs, parameters, and audit trails, improving traceability and "apples-to-apples" comparisons across weekly runs and what-if scenarios.
• Developed a backtesting and accuracy-tracking framework (e.g., MAPE/WAPE, P50/P90 bands) to evaluate forecast quality over time, and inform parameter tuning and model selection.
• Optimized simulation runtime and throughput (e.g., parallelization, partition strategies, caching), and introduced observability (metrics, logs, alerts) to meet strict SLOs for weekly planning cycles.
• Partnered with upstream service owners to define data contracts and SLAs; reduced breakages from schema drifts, and improved on-time availability of inputs for planning deadlines.
• Built stakeholder-facing dashboards/reporting to communicate demand, labor requirements, and sensitivity analyses to Ops, Workforce Planning, and Finance.

Highlighted Project — Production-Simulation Integration (AspectJ/AOP)

• Led the integration of production services with the simulation environment by using AspectJ to intercept service-level API calls and safely inject simulation outputs in place of live data under controlled flags.
• Implemented feature toggles, request/response shims, and parity tests to validate behavior across simulated vs. Production paths enable safe A/B and canary evaluations.
• Reduced integration risk and manual stubbing effort, while allowing teams to test operational changes against realistic, end-to-end simulated conditions before rollout.

• Designed an autonomous quadcopter for post-disaster search & rescue (SAR) to detect humans under debris using thermal + RGB sensor fusion, on-board edge AI compute, and safe coverage patterns.
• Perception: Trained YOLOv5/YOLOv8 person detector (thermal and visible spectra), applied threshold tuning, and NMS, and added DeepSORT/Kalman tracking for multi-target persistence under occlusion.
• Sensors and compute: Integrated FLIR thermal camera, HD RGB, and optional LiDAR/Depth (e.g., RealSense/RPLIDAR) for 3-D mapping; on-board Jetson-class GPU (or Raspberry Pi 4 and Coral TPU) for real-time inference.
• Autonomy & mapping: PX4/ArduPilot flight stack with ROS/ROS 2 + MAVROS/MAVSDK; ORB-SLAM2/RTAB-Map to build an occupancy grid, enabling lawn-mower/cellular-decomposition search and A*/RRT* path planning with geofencing.
• Comms and ground control: Live RTSP video and MAVLink telemetry to QGroundControl; fallback LTE/LoRa link; added TLS/VPN tunnel for encrypted data.
• Simulation and testing: Built rubble scenarios in Gazebo and AirSim for rapid iteration; staged field tests to measure precision, recall, time to first detection, and flight endurance; implemented metric logging with InfluxDB and Grafana.
• Safety and reliability: Configured RTL, obstacle avoidance, prop guards, and no-fly zone limits; fail-safe landing on link loss; battery health, and wind tolerance checks.

• Company Overview: Interned at India's most prestigious research center, Defence Research Development Organisation
• Interned at India's most prestigious research center, Defence Research Development Organisation