I build local-first AI tools, backend systems, and interactive software experiments with a focus on efficiency, modularity, and practical daily-use workflows.
My current areas of interest include lightweight AI deployment, semantic memory systems, speech interfaces, computer vision workflows, asynchronous pipelines, and structured interface implementation.
The general focus is on building scalable, plug-and-play local AI systems that run efficiently on consumer-grade hardware, while still benefiting from higher-end machines for larger models, lower latency, and richer multimodal workflows.
Note: In each repository, you may find a dedicated
/docsdirectory containing case studies, engineering notes, architectural decisions, and assistant rules. These documents are intended to help readers understand the system design trade-offs, follow technical updates, or adapt the assistant guidelines for their own AI projects.
Open To:
- Local-First AI Experiments
- Backend System Prototypes
- Memory & Context Pipeline Research
- Multimodal Interface Workflows
- Open Source Collaboration| Core Tools |
|
|
| Project-Level Systems |
|
| Area | Engineering Focus | Applied In |
|---|---|---|
| Local LLM Systems | Local inference, small-model orchestration, context budgeting, prompt layering | Local assistant runtime, Discord AI assistant, private AI experiments |
| Semantic Memory & RAG | Vector retrieval, metadata filtering, long-term memory design, structured context injection | PostgreSQL + pgvector memory pipelines |
| Speech Interfaces | Speech-to-text, text-to-speech, audio routing, playback queues, sample-rate normalization | Local voice workflow experiments |
| Computer Vision | Image classification, preprocessing, model comparison, moderation pipeline design | TensorFlow / Keras image moderation system |
| Async AI Pipelines | Non-blocking task orchestration, queue-based processing, cancellation-safe runtime design | Python asyncio-based assistant systems |
| Product-Oriented AI Tools | Local-first workflows, modular service boundaries, structured technical interfaces | Workspace and study-tool experiments |
PHASE 1 // Local AI Core & Vector Memory
Local LLM agent gateway with unified PostgreSQL semantic memory mapping.
| Category | Details |
|---|---|
| Stack | Python, Ollama, PostgreSQL, pgvector, Pedalboard DSP, Discord API |
| Scale | Modular local assistant framework designed for consumer-grade hardware |
| Performance | Active scrolling context window, strict token budgeting, async task handling |
| Security | Prompt-layer separation, injection-aware memory handling, local-first architecture |
| Impact | Converts a Discord assistant into a reusable local AI infrastructure layer |
| Repository | AI-Discord-Personal-Assistant |
A local AI assistant framework originally built around Discord, then expanded into a modular backend system for local inference, long-term memory, voice interaction, and asynchronous communication.
The system consolidates relational and vector memory into PostgreSQL with pgvector, reducing architectural fragmentation while keeping the assistant practical for smaller local models.
Key engineering work includes hybrid temporal retrieval, semantic vector memory, active context-window trimming, local TTS routing, audio post-processing, and modular backend separation.
PHASE 2 // Multimodal Orchestration Runtime
Asynchronous local-first conversational runtime with speech, memory, and expressive output control.
| Category | Details |
|---|---|
| Stack | Python Asyncio, Ollama, Faster-Whisper, GPT-SoVITS, Pedalboard DSP |
| Scale | Experimental multimodal client runtime connected to a shared local AI backend |
| Performance | Thread-safe audio queues, resampling pipelines, decoupled LLM and voice synthesis clients |
| Security | Local inference path, modular service boundaries, isolated runtime responsibilities |
| Impact | Establishes the foundation for real-time local AI companion and study assistant systems |
| Repository | Private / Experimental |
This system explores real-time interaction between local language models, speech transcription, speech synthesis, audio playback, and external visual control APIs.
Engineering work includes polyphase audio resampling, concurrency-safe playback queues, dual-client topology, and modular integration layers for local-first multimodal interaction.
Graph-Based Local Memory Workspace
Local-first study workspace with visual graph context routing, note memory, and modular AI workflow design.
| Category | Details |
|---|---|
| Stack | Next.js, TypeScript, IndexedDB, Ollama |
| Scale | Private workspace experiment for notes, chat, graph routing, and local AI study flows |
| Performance | Browser persistence, decoupled modules, local model inference, lightweight interaction model |
| Security | Local-first persistence, reduced external dependency surface, privacy-aware architecture |
| Impact | Turns complex local AI pipelines into a more approachable productivity interface |
| Repository | Private / In Development |
A local-first AI workspace experiment designed to bridge the gap between complex local AI pipelines and non-technical daily-use workflows.
The system explores a dual-view architecture: a conventional chat interface for direct query-response interactions, and a visual node-graph canvas for context routing, model configuration, and modular workflow control.
The architecture is intentionally decoupled so notebook, chat, graph, and optional companion runtimes can evolve independently.
Computer Vision Image Moderation Engine
Real-time image classification gateway using deep-learning CNNs and asynchronous caching.
| Category | Details |
|---|---|
| Stack | Python, TensorFlow, Keras, Python Asyncio, aiomysql, Pillow, Discord API |
| Scale | Discord moderation pipeline with model-backed attachment scanning |
| Performance | TTL cache, async database pool, dynamic input-shape extraction, event-loop safe processing |
| Security | Moderation-oriented classification workflow using safe proxy dataset methodology |
| Impact | Demonstrates practical computer vision integration inside an asynchronous bot infrastructure |
| Repository | CV-ImageModeration |
A computer vision moderation gateway that integrates trained CNN models into an asynchronous Discord bot pipeline.
The project compares custom CNN workflows with pretrained architectures such as MobileNetV2, ResNet50V2, and VGG16 for binary image classification.
The backend dynamically prepares incoming image attachments, normalizes tensors for the selected model, and protects high-volume moderation paths using caching and asyncio-aware locking.
Arduino Game Console
Local 4-player embedded game console prototype with LCD display, buzzer feedback, and physical controls.
| Category | Details |
|---|---|
| Stack | Arduino Nano, 20x4 LCD, I2C LCD Module, Tactile Buttons, Piezo Buzzer, ESP32 Telemetry |
| Scale | Embedded multiplayer game loop with multiple modes and hardware input handling |
| Performance | Non-blocking millis() loop, cooperative software clocks, responsive display and buzzer state machines |
| Security | Local hardware prototype, deterministic input handling, no external network dependency required |
| Impact | Demonstrates low-level timing control, hardware interaction, and embedded gameplay logic |
| Repository | IoT-GameConsole |
A technical embedded systems study focused on building a physical multiplayer console with button input, LCD output, buzzer audio, and multiple reflex-based game modes.
The prototype avoids blocking delays by using cooperative timing logic, enabling responsive gameplay, debounced buttons, screen updates, and buzzer playback to operate together.
Independent Technical Lab
Building private and public software experiments around local AI systems, backend prototypes, semantic memory, speech interfaces, computer vision workflows, and structured interface implementation.
- Designed modular local assistant systems using Python, Ollama, PostgreSQL, pgvector, and asynchronous service layers.
- Built retrieval and context-management pipelines for long-term conversational memory.
- Developed local voice interaction experiments using speech-to-text, text-to-speech, audio routing, and DSP workflows.
- Implemented technical interface experiments using structured layout systems and reusable components.
- Explored consumer-grade AI deployment with efficiency, modularity, and maintainability as core constraints.
Private Development Work
Exploring backend architecture, local-first workflows, and interface implementation through iterative software experiments.
- Built backend-oriented prototypes for practical AI and automation workflows.
- Used Laravel and Python for structured application logic and backend experimentation.
- Used Next.js for private interface work and frontend architecture experiments.
- Focused on clean structure, modularity, and practical system behavior over visual complexity.
Learning:
- Lightweight AI deployment
- Semantic memory systems
- Local inference workflows
- Structured backend architecture
Building:
- Local-first AI tools
- Backend system prototypes
- Graph-based memory workspace experiments
- Speech and multimodal interaction pipelines
Exploring:
- Consumer-grade AI deployment
- Context-management pipelines
- Computer vision workflows
- Technical interface implementation
Open To:
- Local AI Experiments
- Backend Prototyping
- Open Source Collaboration
- Practical Software Tooling
No pattern is perfect, stars aren’t divine.