Work

Fall 2024

Modular VLA Robotics Platform

A simulation-first training pipeline for modular robotic arms using Vision-Language-Action models.

RoboticsNVIDIA IsaacVLA ModelsPython

Objective

To reduce physical training time for modular robot arms by 40% using high-fidelity simulation environments (NVIDIA Isaac) and transferring learned policies via VLA models to physical hardware.

In Retrospect

The simulation-to-reality (Sim2Real) gap was wider than anticipated due to friction modeling inaccuracies. However, the modular architecture allowed us to hot-swap end-effectors without retraining the core vision model.

Lessons Learned

Synthetic data generation is only as good as the physics engine. Investing early in accurate URDF modeling for the modular joints saved weeks of debugging later. VLA models require significant token optimization for real-time inference.

Read Case Study

Summer 2024

Edge AI Security Camera

Next-gen smart home security device featuring on-device person detection and custom silicon integration.

Consumer HardwareEdge AICustom SiliconPrivacy

Objective

Launch a battery-powered security camera capable of running local person detection under 500ms latency without cloud round-trips, extending battery life by 30%.

In Retrospect

Balancing thermal constraints with NPU performance was the critical path. We successfully offloaded audio processing to a low-power DSP, which freed up the main core for vision tasks.

Lessons Learned

Users value privacy (local processing) more than theoretical accuracy improvements. Custom silicon requires a software roadmap 6 months ahead of hardware tape-out to ensure driver maturity.

Read Case Study

Spring 2024

Adaptive Audio Codec Integration

Optimizing real-time voice communication for low-bandwidth environments using AI-driven codecs.

Audio EngineeringCodecMachine LearningC++

Objective

Implement the Lyra/AV1 codec to maintain crystal clear voice quality even at 3kbps packet loss rates, targeting emerging markets with unstable connectivity.

In Retrospect

The initial CPU overhead for the neural decoder was too high for older IoT chipsets. We refactored the inference engine to use integer quantization, dropping CPU usage by 15%.

Lessons Learned

Network conditions in the lab rarely match the 'long tail' of real-world jitter. Always build a fallback to standard codecs (Opus) if the AI model confidence drops.

Read Case Study

Winter 2023

Commercial Grid Storage Dashboard

B2B analytics platform for monitoring GWh-scale energy storage assets.

B2B SaaSEnergyData VizReact

Objective

Provide facility managers with real-time state-of-charge (SoC) visibility and arbitrage opportunities for commercial megapack installations.

In Retrospect

We over-indexed on real-time granularity (second-level data) which bloated the database costs. Customers actually preferred 15-minute aggregates with predictive anomaly detection.

Lessons Learned

In B2B energy, 'trust' is the metric. If the dashboard data conflicts with the utility meter by even 1%, the user churns. Data reconciliation must be the first feature built.

Read Case Study