Alvin Jomon
Software Developer at iKorex engineering full-stack software applications, managing Linux server infrastructure & CI/CD deployments, and building embedded IoT systems.
Software Developer at iKorex & B.Tech CSE student at College of Engineering Poonjar (CEP). Focused on high-velocity development and reliable infrastructure—pairing modern AI-assisted engineering workflows with deep hands-on expertise in Linux server management, shell automation, CI/CD pipelines, and embedded IoT systems.
Core Engineering Highlights
A structured overview of software engineering roles, capstone architectures, campus leadership, and technical specializations.
Agent Jarvis, Project Doremon & Central Brain
Private on-premises voice AI, assistant orchestrator, and local vector memory ecosystem. Features sub-500ms CUDA speech recognition via Distil-Whisper, PipeWire neural Silero VAD, hybrid 1024-dim vector RAG with SQLite FTS5 BM25 search, and zero cloud data leaks.
Software Developer @ iKorex
Building full-stack software applications, managing production Linux servers, writing custom shell automation, and coordinating team tasks for fast, dependable releases.
TinkerHub, FOSS Club & IEDC Execom
Serving as Campus Learning Coordinator on the TinkerHub Core Team, leading open-source projects in FOSS Club CEP, and driving startup initiatives as an IEDC Executive Member.
Video Editing, Poster Design & Shopify
Hands-on freelancing experience providing high-end video editing in DaVinci Resolve, graphic poster designing in Adobe & Penpot, and end-to-end e-commerce store setups.
Flagship Projects Showcase
Detailed breakdown of key software architectures, AI memory engines, hardware setups, and systems daemons.
Agent Jarvis (Voice AI & Orchestrator)
Autonomous, 100% private local voice assistant and terminal orchestrator. Features noise-immune neural Silero VAD, CUDA-accelerated Distil-Whisper STT (INT8), local Ollama planning, and a cyberpunk HUD.
- ✓ Neural Silero VAD & Transient Detection: 16 kHz PipeWire stream with dual-clap triggers and noise immunity.
- ✓ CUDA GPU Whisper STT: Distil-Whisper INT8 inference running on CUDA in ~480ms with 266MB VRAM footprint.
- ✓ Working Video Demo Included: Live demo showing neural speech recognition, LLM planning, and terminal actions.
# Neural VAD & Intent Trigger (trigger_detector.py) vad_prob = silero_vad(pipewire_pcm) if vad_prob > 0.5: transcription = cuda_whisper(audio_chunk)
Project Doremon
Local, privacy-first Home Assistant AI agent with dynamic emotional state tracking (battery, energy, caring), role-based secrecy filters, and hybrid RAG memory.
- ✓ Role-Based Secrecy Filters: Secrets tagged `"don't tell Dad"` are strictly filtered at database layer before LLM context generation.
- ✓ Hybrid Vector + BM25 RAG: Fuses `mxbai-embed-large` 1024-dim vectors with SQLite FTS5 BM25 search via Reciprocal Rank Fusion.
- ✓ Dynamic Emotional State Engine: Simulates mood shifts (Joyful, Concerned, Tired, Helpful) based on user conversations.
# Doremon Database Secrecy Filter (db_manager.py) if memory.get("confidential_from") and current_speaker in memory["confidential_from"]: continue # Filtered out before reaching LLM context! results = reciprocal_rank_fusion(vector_matches, bm25_matches)
Project Happy
Smart Home Voice Assistant ecosystem featuring a Kotlin/Jetpack Compose Android companion app, offline Vosk STT, Kokoro TTS, and ESP32 microcontroller clients.
- ✓ Offline Voice Engine: Instant Vosk speech-to-text & Kokoro text-to-speech without cloud dependency.
- ✓ UDP Beacon Discovery: Auto-discovers local laptop server on Wi-Fi network over port 9990.
- ✓ On-Device Gemini Fallback: Runs Gemini AI locally on smartphone when laptop hub is offline.
// UDP Auto-Discovery Protocol (Android & ESP32) val socket = DatagramSocket(9990) val packet = DatagramPacket(buffer, buffer.size, broadcastAddr, 9990) socket.send(packet) // Discovers local server URL automatically
Central Brain Engine
Local-first, zero-cloud vector memory retrieval CLI tool combining 1024-dimension embeddings (`mxbai-embed-large`) with SQLite FTS5 BM25 search.
- ✓ CLI Knowledge Engine: Instant retrieval via `brain query "
"` tool in terminal. - ✓ Dual Hybrid Search: Blends dense semantic vector distance with sparse BM25 keyword frequency.
- ✓ Document Ingestion & Pruning: Indexes specs and markdown documentation into central memory.
# Local CLI Central Brain Command ~/.local/bin/brain query "Linux kernel driver fix" ~/.local/bin/brain remember "Fixed sleep resume bug" --category "Fix"
ASUS TUF Backlight Ambience Daemon
Lightweight ambient lighting daemon for ASUS TUF laptops syncing keyboard RGB backlight colors dynamically with active wallpapers or real-time PipeWire screen video.
- ✓ PipeWire Screen Ambience: Captures screen color output in real-time via PipeWire & GStreamer.
- ✓ Saturation-Squared Algorithm: Highlights vivid wallpaper colors while filtering out dark noise.
- ✓ Flicker-Free Transitions: Dynamic alpha adaptation ensures smooth LED color morphing.
# Saturation-Squared Color Extraction (plasma-video-ambience.py) weighted_val = lightness * (saturation ** 2) target_alpha = 0.08 if calm_scene else 0.85 # Smooth adaptation