Work & Research

Overview ------------------ This project concerned training and deploying a small CNN to fit on a mere 256kb memory buffer. The model was then deployed on a real edge-hardware device, where it was used for wakeword detection for a seperate school project. Technical Summary ------------------ -The neural network was trained entirely in Google Collab, primarily using Tensorflow(specifically TFLite Micro). -It was designed to fit on a RAM buffer of only 256kb in order to accomodate the Arduino Nano 33 BLE Sense microcontroller. -The model was trained from scratch on a wakeword dataset, enabling it to detect "open" and "close" as wakewords. -The final model fit into 200-256kb of RAM, and had roughly that many parameters due to it's native FP8 quantization level. Takeaways ------------------ -Through this experience, I gained experience with a variety of AI/ML tools, including: -Google Collab with Python -Tensorflow/TFLite/TFLite Micro -Arduino IDE I also gained crucial foundational experience with Neural Networks, experimenting with such concepts as optimization, loss, gradient descent, layering, and activation functions.

TECHNICAL OVERVIEW ------------------ BasecampAgent V1 was developed to solve the latency overhead found in modern AI frameworks. By utilizing direct LLAMA.CPP calls and raw Python, the system achieves direct communication with the hardware, specifically optimized for the STRIX HALO architecture. MULTI-TIER HARDWARE ORCHESTRATION --------------------------------- The architecture dynamically offloads tasks between disparate hardware units to balance speed and reasoning depth: // ORCHESTRATOR (AMD RX 9070) Handles fast-logic routing, the primary chat interface, and immediate tool execution. // HEAVY AGENT (STRIX HALO) Utilizes 128GB Unified Memory for deep-research tasks and long-form reasoning chains that would overflow standard VRAM buffers. HARDWARE-LOCKED TOOLING ----------------------- To maintain system stability on shared memory architectures, a MUTEX-BASED LOCKING SYSTEM was implemented. This ensures that the iGPU is not overloaded by simultaneous requests from different sub-agents. KEY FEATURES ------------ - SMART_SEARCH: Parallelized web scraping and synthesis. - DEEP_THINK: Delegated long-form reasoning. - EXECUTE_COMMAND: Direct Ubuntu bash integration. - STATE-AWARE INJECTION: Hardware status (Busy/Available) is fed into every prompt. PERFORMANCE DATA ---------------- RESULT: 35% REDUCTION IN TTFT Measured against standard LangChain-based local implementations. NOTES ----- While V1 proved the efficacy of "No-Framework" agency, the roadmap for V2 involves moving toward a more modular system to balance performance with ease-of-use.