Kavin Vasudevan

Architected a FreeRTOS-based firmware for an electric vehicle's custom Battery Management System, coordinating cell monitoring, charging, fault detection, and safety-critical state transitions across concurrent tasks.

• Safety-critical state machine for charging, balancing, and fault handling
• Concurrency-safe data sharing across monitoring, telemetry, and control tasks
• CAN telemetry and fault reporting to vehicle subsystems

Developed STM32 firmware for the Powertrain Control Unit, processing accelerator and brake inputs via ADC with normalization, plausibility checks, and fault detection, then transmitting values over CAN for safe, real-time vehicle control.

• Real-time pedal and brake sensor processing
• Normalization + plausibility checks for safety
• Millisecond-interval CAN telemetry for closed-loop control

Designed the dashboard firmware with I²C I/O handling, CAN decoding, and real-time UI updates, adding fault indicators (IMD, BMS) and a Ready-to-Drive buzzer synchronized with the BMS state machine.

• Superloop architecture with 10 ms refresh cycle
• Handles 3 driver switches + 2 safety buttons via I²C expanders
• Displays battery status, speed, tractive system, and coolant system data

Implemented a circular buffer for high-throughput CAN logging, ensuring lossless real-time telemetry storage with minimal memory usage and providing persistent data for post-race analysis and debugging.

• Circular buffer for high-throughput CAN capture
• Persistent logging to SD card
• Live wireless telemetry streaming via XBee

Extended Pintos OS by implementing a priority scheduler, kernel-mode synchronization primitives (locks, semaphores, alarm clock), and a UNIX-style file system with caching and hierarchical directories.

• Priority scheduler with donation to prevent inversion
• Kernel/user mode isolation with safe system calls
• FAT-style file system with caching, extensible files, and directories

Developed a secure file-sharing platform in Go with user authentication, AES/RSA encryption, and revocable access control, enabling safe collaboration across multiple users.

• Linked-list file blocks with HMAC for integrity and secure appends
• Deterministic UUID-based key generation for encryption
• Fine-grained file sharing with re-keying + revocation

Fine-tuned a BERT model to classify corporate sustainability statements as genuine or greenwashing, achieving 99% accuracy and highlighting linguistic markers that distinguish authentic impact from vague claims.

• Fine-tuned BERT for sustainability statement classification
• 99% test accuracy with CI (0.97–1.00)
• Detects linguistic markers of genuine impact vs greenwashing

Built a machine learning pipeline on FIA/WEC lap data, engineering features like start position and speed efficiency; trained Random Forest models achieving ~70% accuracy and an AUC of 0.82.

• Feature engineering on FIA/WEC lap telemetry
• Random Forest model with ~70% accuracy, AUC 0.82
• Identified start position, speed efficiency, and sector times as key predictors

Implemented a nanoGPT-style Transformer in PyTorch to evaluate in-context morphological learning, training a 10M-parameter model that achieved 63.5% accuracy on validation tasks across 18 transformations.

• Custom 10M-parameter Transformer for morphological learning
• Evaluation across 18 transformations with 63.5% accuracy
• Benchmarks in-context learning limits on irregular forms