Screen Watcher
Local desktop application that detects visual changes and builds a semantic history of screen activity.
VIEW REPOSITORY ↗LOADING / NEXT VIEW
Preparing the next section…03 / CHALLENGE COMPLETE
Thirty days of daily development. From independent experiments to a commercial product, ending where the journey comes together: this portfolio.
01—09 / FOUNDATIONS
Local desktop application that detects visual changes and builds a semantic history of screen activity.
VIEW REPOSITORY ↗Key-value database built from scratch to explore indexing, append-only storage and log compaction through benchmarks.
VIEW REPOSITORY ↗Real-time market data pipeline that streams Binance trades through Kafka to independent consumers and a live dashboard.
VIEW REPOSITORY ↗One-day experiment comparing Unity and Unreal Engine by building the same Android face-tracking AR experience.
VIEW REPOSITORY ↗Local RAG system with structure-aware retrieval, grounded generation and traceable sources over academic documents.
VIEW REPOSITORY ↗Adaptive vision system where global Cloud detections guide local Edge attention, reducing remote inference calls by 71% and transferred data by 74% in a reproducible video benchmark.
VIEW REPOSITORY ↗From-scratch web search engine with controlled crawling, an inverted index and BM25 ranking, revisiting a personal search engine project from 2022.
VIEW REPOSITORY ↗Browser performance experiment comparing main-thread and Web Worker GLB loading, measuring responsiveness, blocking time and time to first rendered frame across increasingly large 3D assets.
VIEW REPOSITORY ↗Persistent Windows desktop pet with global eye tracking, direct interaction and a lightweight Tamagotchi-style state that continues evolving while the application is closed.
VIEW REPOSITORY ↗11—19 / NEXT ITERATION
Native Android 17 Settings extension for inspecting framework-reported cellular information, including serving and observed cells, network identity, radio technology, frequency and signal metrics. Built directly into AOSP and validated on Cuttlefish.
VIEW REPOSITORY ↗Autonomous black-box web QA system combining deterministic browser checks, LLM-driven semantic exploration and VLM screenshot perception, with validated execution, replay and evidence-based defect reporting.
VIEW REPOSITORY ↗Custom CUDA Gaussian Blur benchmark comparing naive and shared-memory tiled kernels against OpenCV CPU and OpenCV CUDA. Separates kernel-only from end-to-end performance to expose the impact of GPU memory transfers.
VIEW REPOSITORY ↗Visual cryptography benchmark measuring real CPU and RTX 3090 CUDA SHA-256 throughput, then extrapolating exhaustive-search times across increasingly large key spaces. Includes matched CPU/GPU comparison, PBKDF2 and controlled synthetic secret recovery.
VIEW REPOSITORY ↗Hardware-in-the-loop Pong experiment comparing FPGA and STM32 controllers under the same Windows-hosted physics and serial protocol. Implements a quantized neural-network policy in SystemVerilog on an Arty A7-35T and a FreeRTOS controller on STM32, with real-time dual-UART telemetry and physical hardware execution.
VIEW REPOSITORY ↗Experimental Android benchmark exploring how far self-hosted push notifications can work without Firebase or another push provider. Compares WebSocket, MQTT, periodic polling and foreground-service delivery across background, screen-off, Doze, network-loss and process-death conditions on a physical Android device.
VIEW REPOSITORY ↗Quantization benchmark mapping the speed, efficiency and quality trade-off of Qwen3-8B from F16 down to Q2_K on an RTX 3090. Measures throughput, VRAM, energy, perplexity, KL divergence and GSM8K accuracy, identifying Q4_K_M as the compression knee.
VIEW REPOSITORY ↗On-device LLM experiment measuring how large a model a physical Android phone can run before it stops being interactive. Runs Qwen3 0.6B–8B (Q4_K_M) through llama.cpp via NDK/JNI on a Galaxy Z Fold6 CPU, fully offline, recording load time, prefill/decode throughput, memory and thermal throttling to separate fits, usable and interactive.
VIEW REPOSITORY ↗Local zero-shot voice-cloning benchmark measuring how much reference audio a cloner needs to sound like the speaker. Climbs a 3 s → 5 min reference ladder for XTTS-v2, Chatterbox Multilingual and F5-TTS (Spanish) on an RTX 3090 using Multilingual LibriSpeech readers, scoring ECAPA speaker similarity, Whisper ΔWER, DNSMOS, RTF and VRAM. Finds that raw audio plateaus after ~10 s, while picking the best 10 s window from 5 min lifts similarity to 0.775 (ceiling 0.904).
VIEW REPOSITORY ↗21—29 / ONE COMMERCIAL PRODUCT
Nine days focused on one commercial product. After the independent experiments of Levels I and II, the final stretch brought that daily practice into a single, sustained build.
Days #21–#29 belong to the same product: nine days of focused development.
The starting point of Level III: a connected 3D desktop companion evolving into an independent product. After a week of development, the system already combines a native transparent Windows experience, autonomous pet behaviors, a web client and a FastAPI/PostgreSQL cloud backend. Shared pet identity, persistent state, exclusive presence and cross-device synchronization are implemented and tested. The next ten milestones focus on turning this technical foundation into a product ready for real users.
Preparing the product for its first users. Built an integrated analytics dashboard to monitor registrations, pet adoptions and returning users, providing visibility into early adoption and engagement. Also produced a 20-second promotional video, the first potential advertising asset. The focus is shifting from technical development toward launch readiness, product visibility and understanding real-world usage.
Launched the product's first promotional campaign on TikTok, moving beyond development into audience acquisition. The campaign introduces the digital companion to potential users and establishes a first opportunity to observe reach, engagement and interest outside the development environment. An initial step toward understanding whether the product can attract an audience.
Refined the product landing page based on early insights from the first advertising campaigns. Initial traffic revealed a predominantly mobile audience and potential confusion between browser access, the Windows desktop experience and the upcoming Android app. Introduced subtle changes to clarify platform availability, improve the messaging hierarchy and make the web experience easier to discover. A first conversion-focused iteration driven by real visitor behavior.
Set up a visual multi-agent development office using Clawn3D to coordinate work across the product ecosystem. Started with a focused PetWeb task to evaluate the agents' capabilities and, after a successful initial result, expanded their responsibilities to the full set of product subprojects. An experiment in delegating development workflows, coordinating parallel work and moving toward agent-assisted product management.
Reviewed the results of the first paid acquisition experiments. TikTok generated visibility, while Meta Ads successfully drove traffic to the landing page, but neither produced measurable registrations or pet adoptions. The findings raised questions about audience targeting, platform positioning and the conversion journey. A reality check on the challenges of turning a working product into one that attracts users.
Extended an existing Claw3D-based visual agent office with small improvements and tested its use across the product's development workflows. An initial focused task succeeded, but broader delegation revealed interruptions, frequent supervision needs and limited parallel execution. With only minimal tuning, the experiment was paused rather than fully optimized. An early exploration of agent-assisted development, highlighting both its potential and the challenges of practical orchestration.
Revisited the engineering work completed before Stage III began. Over ten days of development, a virtual companion evolved into a connected product combining a Unity-based Windows client, a browser experience and a FastAPI/PostgreSQL cloud backend. Autonomous behaviors, persistent identity and cross-device synchronization formed the technical foundation for its public launch. A look behind the scenes at the development effort that preceded the product milestones.
Revealing Krovyte, the independent digital companion developed throughout Stage III. Built with Unity, C#, FastAPI and PostgreSQL, Krovyte combines an autonomous Windows desktop pet, a browser-based experience and persistent cloud synchronization. After its first public launch and advertising experiments, the product is now presented as part of the 30 Projects journey. Its source code remains private, while the product itself is publicly accessible.