I'm an engineer and developer with a keen focus on embedded systems, avionics and artificial intelligence. My interests span firmware development and hardware integration to intelligent applications and comprehensive software solutions. I’m driven by a curiosity to understand complex systems, a passion for exploring emerging technologies, and the challenge of transforming ideas into robust, efficient, and meaningful solutions.
Four systems, four different problems: a company knowledge graph, a cross-device music player, a quantitative research engine, and a voice assistant.
Neura1 · OpenWave · Agent Alpha · Captain AI
A company brain that shows what a change will affect.
Neura1 connects a manufacturer's parts, builds, customer orders, people, and decisions in one knowledge graph. When something changes, it traces the impact, explains why, and helps the team plan what to do next.
What the system brings together
- Knowledge with evidence. Decisions, documents, conversations, and people are linked to the underlying work. An answer can lead back to the source that supports it.
- Change impact. A change to a part or supplier can be traced through product revisions, planned builds, customer orders, and the people responsible for them.
- Planning before promises. A timeline compares the original plan with the current forecast; stock, lead times, orders, and availability explain where a delivery may slip.
- Action in existing workflows. Department agents surface issues, while Zoho connections and MCP tools let Neura1 work alongside the systems a team already uses.
Inside the engineering — graph, planning, and trustworthy AI
FalkorDB joins company records into a graph of parts, products, builds, suppliers, orders, customers, people, and decisions. Traversals reveal dependencies that would otherwise sit across separate spreadsheets and tools. Source links keep the reasoning inspectable.
The planning engine works from inventory, open purchase orders, supplier notices, lead times, and leave to forecast delays and explain their causes. A sandbox can test a proposed change on a copy of the graph before it affects the working plan. Cost and stock calculations stay in code rather than being guessed by a model.
The model helps find and summarize evidence, while validation checks quotes against sources and applies access rules to answers. A web interface, chat, document tools, and optional MCP connections expose the same underlying system to different teams. The repository includes a product demonstration and evaluation notes.
PYTHON / FALKORDB / NEXT.JS / MCP / LOCAL INFERENCE
Explore the code ↗ · Watch the demos · Read the evaluation
One music library. Every device.
OpenWave is an open-source music player designed to feel like one product across every screen. It brings gapless playback, offline downloads, synced libraries, and listen-together sessions to desktop, mobile, web, and self-hosted setups.
What the system brings together
- One player across devices. A shared Python engine powers macOS, Windows, Linux, and Android apps, an iPhone web app, and a Docker deployment.
- The listening experience. Gapless playback, synced lyrics, offline downloads, personal mixes, and playlist import turn a catalogue into a daily-use music library.
- Your library, in your hands. Profiles, playlists, likes, and history live in SQLite on your device or server, with changes synced across your devices.
- Listening together. Devices on the same network can discover and pair with one another, hand off playback, act as remotes, or join a synchronized session.
Inside the engineering — playback, sync, and release delivery
FastAPI provides the catalogue, streaming, library, import, and device APIs. A shared web interface runs inside platform-specific shells, including Swift on macOS and Kotlin on Android. The same core behaviour reaches desktop, mobile, browser, and home-server use cases.
The engine resolves audio tracks and proxies streams in chunks so playback can resume after an interrupted connection. Library operations are stored locally in SQLite; clients keep an outbox of changes made while offline and sync on reconnection. Nearby devices require pairing before they can control playback or join a session.
The project includes installers, Docker images, checksums, in-app updates, screenshots, and setup guides. The release path matters as much as the player itself: someone should be able to install it and use it without reconstructing the development environment.
PYTHON / FASTAPI / JAVASCRIPT / SQLITE / DOCKER
Explore the code ↗ · Visit the website · Download
From market data to signals, risk decisions, and a paper-trading arena.
Agent Alpha is my quantitative research and trading project for Indian equities. It brings market-data ingestion, technical analysis, machine-learning signals, market-regime classification, risk checks, and portfolio monitoring into a connected system. The engineering spans the analytical backend, scheduled jobs, persistent data, desktop interface, and remote monitoring.
What the system brings together
- Multiple views of the market. Technical indicators, momentum, machine-learning predictions, institutional flows, options data, and statistical-arbitrage research contribute to the analysis. Ensemble weighting combines signals into a directional assessment.
- Market context and risk controls. A four-state Hidden Markov Model classifies market conditions. A separate veto layer can override a signal, while volatility and Kelly-based sizing inform position allocation.
- Research and simulated execution. Screening, individual-stock analysis, portfolio views, and a paper-trading arena connect research with simulated positions and outcomes. Backtesting and signal-accuracy views support evaluation of the strategies.
- An operational interface. The dashboard is accompanied by an Electron desktop wrapper, a macOS menu-bar utility, and Telegram commands and alerts. These provide different ways to inspect the system and monitor its activity.
Inside the engineering — models, infrastructure, and reliability
The documented model families cover trend and volatility indicators, momentum, boosted-tree classifiers, sequence models, institutional flows, and pairs research. The regime classifier adds context to those signals: an assessment in a quiet uptrend should be treated differently from one during a volatile drawdown. The risk layer then considers whether a candidate should proceed at all.
The deployment design separates a Mac-based interface and development workstation from an always-on Windows execution host. Docker services, scheduled scans, a FastAPI backend, PostgreSQL persistence, and Ollama-based language-model analysis support the system. SSH deployment tooling connects the development and execution environments.
The repository includes checks for database hangs, portfolio queries, pricing fallbacks, regime logic, statistical calculations, and the paper-trading engine. This matters because a research system depends on both its analytical logic and the quality and availability of the data feeding it.
PYTHON / FASTAPI / POSTGRESQL / OLLAMA / ELECTRON / DOCKER
Explore the code ↗ · Architecture & research · Interface gallery
A voice interface connected to the screen, the desktop, and the phone.
Captain AI is my modular assistant project, built around a central intent router that coordinates speech, conversation, screen understanding, and device actions. It combines on-device audio processing with cloud language-model inference and local vision, bringing those capabilities into a desktop overlay and a remote Telegram interface.
What the system brings together
- A complete voice loop. OpenWakeWord detects the trigger, Faster-Whisper transcribes speech, and Piper generates spoken responses. Audio ducking lowers music during an interaction; cached speech output reduces repeated synthesis work.
- Conversation with retrieved context. Groq supplies streaming language-model responses, with Wikipedia and DuckDuckGo retrieval providing context for information requests. Ollama and Moondream handle local screen-image analysis.
- Actions across devices. Intent handlers connect voice commands to application launching, web navigation, email workflows, music, and Android automation through ADB. Persistent memory, to-do lists, alarms, and timers support everyday tasks.
- A visible assistant state. The Ghost HUD uses a frameless PyQt6 window and an HTML/CSS interface. WebSocket messages connect it to the backend, displaying listening, processing, and speaking states. A hotkey summons the overlay; Telegram provides another command channel.
Inside the engineering — orchestration, audio, and the Ghost HUD
The central Brain routes input to independent speech, language, vision, and device-control engines. The architecture keeps microphone capture, model calls, audio playback, and interface updates as distinct responsibilities connected through the orchestration layer.
The voice pipeline includes filters for spurious transcription, wake-word sensitivity, speech caching, and coordination between music and spoken output. Wake-word detection, speech recognition, speech synthesis, and the Moondream vision path run locally. The documented default conversational model uses Groq; retrieval and messaging integrations require network access. Android integration uses ADB commands and UI hierarchy parsing, and the repository demonstrates voice-driven email composition and phone workflows.
The project progressed from a CustomTkinter prototype with logs and controls to the Ghost HUD: a compact overlay with a global shortcut and visual feedback for each interaction stage. A WebSocket protocol carries state changes between the Python backend and the interface. The repository includes screenshots of both generations and video demonstrations. Windows is the documented setup target; broader platform support remains development work.
PYTHON / PYQT6 / WEBSOCKETS / GROQ / OLLAMA / WHISPER / PIPER / ADB
Explore the code ↗ · Architecture & demonstrations · Watch the Gmail workflow
Embedded systems & avionics
The intersection of hardware, software, and dependable system behaviour.
Applied AI & knowledge systems
Local inference, voice interfaces, graph-backed reasoning, and agents connected to useful tools.
Product engineering
Taking an idea through interface design, backend services, packaging, documentation, and releases.
Applications Python · TypeScript · JavaScript · React · Next.js · Flutter
Services & data FastAPI · SQLite · PostgreSQL · FalkorDB
AI & integration Ollama · vLLM · Whisper · MCP
Build & delivery Git · GitHub Actions · Docker · Cross-platform packaging
A few things I care about when building
- Interfaces that make the next action obvious.
- Data ownership and practical self-hosting options.
- Traceable outputs and clear limits on what a system knows.
- Documentation and releases that make a project usable beyond my own machine.
Interested in embedded systems, applied AI, or building useful open-source software?
Start with an issue on the relevant project: OpenWave · Neura1 · Agent Alpha · Captain AI.
SAI SIDDHARTH B S / EXPLORING SYSTEMS. BUILDING PRODUCTS.