Chen Ping

Led design of financial-grade frontend observability platform (ORS), supporting real-time collection, storage and sub-second analysis of 300-400M daily frontend tracking data (including page performance metrics, JS errors, user behavior trajectories); addressed original architecture's high query latency (5s+), data skew and write bottleneck issues by refactoring to Kafka + StarRocks streaming-batch integrated architecture: achieved high-concurrency data buffering and peak shaving through Kafka cluster, built distributed MPP OLAP analysis layer based on StarRocks; designed layered and bucketed storage strategy (time-based partitioning + app ID hash bucketing), combined with vectorized execution engine and materialized views, achieving low-latency multi-dimensional analysis at PB scale, ensuring full-link monitoring and second-level fault location for financial business.

Designed and implemented multimodal emotional venting interaction system (complementary psychological intervention product matrix to hugging robot): Built "stimulus-feedback-evaluation" real-time closed-loop architecture, integrating pressure sensor array (punching force detection), photoplethysmography (PPG) heart rate/blood oxygen collection, forming physiological-behavioral dual-channel perception system; designed dynamic interaction logic based on reinforcement learning strategy engine: large model adjusts feedback strategy in real-time based on user punching force and physiological stress indicators (heart rate variability HRV, blood oxygen fluctuation) — within safety threshold, LLM generates progressive provocative context to stimulate venting, when physiological indicators exceed limits (HR>140 or SpO2<95%) or punching force exceeds safety boundary, automatically switches to pleading/comfort mode and reduces interaction intensity; system transforms psychological venting into quantifiable physiological data monitoring through controllable physical interaction, providing objective evaluation means for emotional management.

Led design and implementation of high-concurrency breeding IoT management platform (2022-2025, as R&D Manager), supporting real-time access and control of 40K breeding environment monitoring devices (temperature, humidity, ventilation); addressing breeding industry's pain points of budget sensitivity, poor network conditions, and need for private deployment, adopted "lightweight monolithic architecture + edge disaster recovery" strategy (Spring Boot + Redis + MongoDB + MySQL), rejecting microservices拆分 to reduce deployment complexity and resource consumption; through TCP Modbus communication layer optimization (connection pool reuse, asynchronous polling scheduling), achieved 30ms-level remote control and parameter delivery supporting 40K devices on single machine; designed tiered alarm engine (device power failure/environmental anomaly to SMS/phone/APP tiered push), ensuring second-level perception and emergency response to breeding house environmental anomalies; platform supports one-click private deployment (Docker Compose orchestration), customers can independently complete full platform setup on local servers (8GB memory level), meeting agricultural enterprises' compliance requirements for data not leaving the premises.

Led edge-cloud collaborative architecture refactoring for psychological counseling robot (industry-university-research project during Master's degree after 13 years of work, through two generations of technical evolution): First generation architecture (validation phase): Based on RK3588+Ubuntu designed local full offline solution (local ASR + cloud LLM + local TTS), quickly validated product feasibility, but faced poor TTS effect (synthetic voice mechanical feeling), excessive hardware cost (RK3588 single board cost > CNY 1,500), 35W power consumption and heat dissipation issues; Second generation architecture (mass production phase): Identified "voice quality" and "hardware cost" as scaling bottlenecks, decisively switched to ESP32 + cloud-native architecture: Edge: Based on open-source Xiaozhi voice protocol, only保留 voice collection, tactile sensor driving and streaming playback, hardware cost reduced to < CNY 80 (90% reduction), power consumption reduced to 5W, supporting battery life > 48 hours; Cloud: Used LLaMA-Factory to domain-fine-tune Qwen-32B based on open-source psychological datasets (SFT+RLHF), solving general model's hallucination and emotional understanding deficiencies in psychological counseling scenarios; Built high-concurrency inference cluster based on vLLM on 4xRTX 4090 servers, supporting multi-path streaming session concurrency; Experience upgrade: Connected to cloud Doubao-TTS, voice naturalness MOS score improved from 3.2 to 4.5, supporting multi-emotion voice dynamic switching; Core innovation: Designed low-power tactile-voice linkage protocol, ESP32 initiates context-aware emotional comfort sessions to cloud LLM through WebSocket/UDP+MQTT multi-channel redundancy, achieving <1s closed loop of "physical touch to cloud emotional computing to voice feedback".

Designed and implemented immersive VR spinning bike system (constituting physiological signal product matrix with massage chair and brainwave game projects, developed simultaneously): Adopted Bluetooth bidirectional communication architecture, real-time collection of user riding data (cadence, speed, heart rate), achieving audio-visual multimodal feedback through VR scene synchronization engine: Positive feedback: Dynamically adjusted VR video playback rate based on riding speed, designed C++ native playback engine (based on FFmpeg/GLSurfaceView) replacing Android MediaPlayer, solving high-bitrate 4K video dynamic compensation stuttering issues, achieving smooth 60FPS playback; Reverse control: Built scene-hardware linkage protocol, parsed VR video slope metadata, real-time adjustment of bike electromagnetic resistance through Bluetooth GATT commands (uphill + resistance / downhill - resistance), creating realistic riding experience; System breaks traditional spinning bike's "human passively follows machine" mode through bidirectional data closed loop, achieving real-time bidirectional interaction of "video content to user physical sensation", improving aerobic training immersion and fun.

Designed and implemented BCI-based neurofeedback training system (sharing core architecture with massage chair project): Adopted multimodal physiological signal fusion architecture, real-time collection of EEG, relaxation, and attention three-channel data through Bluetooth BLE, built real-time signal processing Pipeline (filtering to feature extraction to threshold determination to game state mapping); innovatively designed three adaptive neurofeedback games, transforming abstract neural signals into visualized game logic: Ant Maze: State machine-driven based on attention threshold, achieving "high focus orderly forward / low focus random wandering" behavioral feedback; Tortoise and Hare: Adopted sustained attention maintenance mechanism, controlling character movement continuity through attention concentration, training sustained focus ability; Precision Shooting: Designed attention stability determination algorithm (sliding window variance calculation), achieving "higher focus to more stable crosshair to higher shooting accuracy" progressive feedback, supporting manual/auto dual-mode shooting; System transforms neuroplasticity training into real-time visual feedback through gamification closed loop, helping users intuitively perceive and train their attention regulation ability.

Led design of edge-cloud collaborative architecture for mental health intervention (industry-university-research project during Master's degree after 13 years of work): Android端 through multi-channel (Bluetooth/USB) real-time collection of blood oxygen, heart rate and other physiological signals, after edge-side preprocessing联动 with cloud algorithm models, real-time recognition of user anxiety/relaxation state; based on state machine engine driving music recommendation and massage chair and other multi-device collaboration, achieving <200ms latency multimodal mind-body intervention closed loop.

Designed and implemented Android architecture framework for smart breeding, addressing pain points of weak network coverage, high-concurrency offline operations, and cross-end development resource shortage in pig farm on-site management; borrowed Spring Boot's dependency injection and auto-configuration concepts, self-developed lightweight IoC container, built asynchronous message bus based on producer-consumer pattern (replacing traditional Handler communication), unified encapsulation of Wi-Fi/LAN port/RS485 type Modbus hardware access layer through abstract factory pattern; framework achieved "backend development mode前移": adopting MVC layered architecture + annotation-based dependency management, enabling Java backend developers to develop business modules without mastering Android lifecycle details, single person can simultaneously undertake backend API and mobile function development, significantly reducing labor costs; for pig farm multi-building, multi-batch management scenarios, designed plugin-based module architecture, supporting hot-pluggable and independent updates of functional modules (feeding, environmental control, inventory), adapting to customized needs of different scale pig farms.

Led comprehensive technical refactoring of smart key IoT platform (.NET to Java SSM), addressing original system's performance bottlenecks, vendor lock-in, and poor scalability pain points, supporting QR code + acoustic dual-mode door opening smart access control scenarios; for visitor temporary authorization and near-field no-network scenarios, self-developed acoustic near-field communication protocol (based on Manchester encoding, cross Android/iOS dual-end audio encoding/decoding, hardware-end DSP decoding), achieving offline secure identity transmission (distance <50cm, anti-recording replay attack); built Socket long-connection gateway supporting real-time door opening commands and device status synchronization; designed WeChat ecosystem multi-end unified architecture (mini-program + official account service account), achieving tiered permission management and unified account system for owner long-term authorization and visitor temporary QR code.

Led design of smartphone infrared remote control ecosystem (early IoT project 10 years ago), addressing pain points of traditional universal remote's large data volume and poor cross-brand compatibility; innovatively implemented audio-to-infrared conversion architecture: utilizing phone 3.5mm headphone jack dual-channel叠加 38kHz carrier (NEC standard carrier), driving infrared diode to achieve "acoustic to optical signal" hardware-level control, enabling ordinary smartphones to control home appliances without dedicated infrared modules; for complex infrared protocols such as air conditioners, self-developed encoding/decoding algorithm engine, upgraded from original "raw waveform recording-playback" mode to standardized protocol parsing (address code + command code + inverse code verification), achieving remote control data volume compression from several KB level to 4-byte level; built cloud remote control code library platform, supporting user community contribution and sharing.