Developer Needed: Private AI Dictation App for Windows and Android and Mac

I’m looking for a developer to build or customise a private AI dictation platform similar in everyday use to Wispr Flow, for Windows, Mac and Android. Using suitable open-source software is welcome. Please propose the architecture, AI models and implementation approach you believe best meet the requirements, with a fixed-price quote by milestone. PROJECT GOAL Users should be able to sign in and dictate into applications on their computer or phone. Speech should become polished text with minimal delay. All transcription, AI text cleanup and translation must run on our own Proxmox server. No third-party cloud AI processing is permitted. EXISTING SERVER Proxmox host. AMD Ryzen 5 9600X, 6 cores / 12 threads. 128 GB DDR5 RAM. 2 TB NVMe SSD. No dedicated GPU currently confirmed. Other workloads already run on this server. Assess available resources and deploy the backend in a dedicated Linux VM. Recommend hardware upgrades only where justified by testing. WINDOWS AND MAC APPS Run in the system tray or menu bar. Configurable push-to-talk and start/stop shortcuts. Capture microphone audio and connect securely to our server. Insert processed text into the active application. Display recording, processing and connection status. Provide a clipboard fallback where direct insertion is unavailable. Include straightforward installation and permission setup. ANDROID APP Custom voice keyboard for dictating into WhatsApp, Gmail and other applications. Secure login and simple device pairing. Send audio to our server and insert returned text into the selected field. Support various Android phones without requiring local AI models. Clearly handle interrupted connections and server unavailability. A web recording page alone is not sufficient for system-wide dictation. MULTIPLE USERS AND ADMINISTRATION Individual accounts and multiple devices per user. Synchronised personal dictionaries, language preferences and writing styles. Optional private transcript history with deletion controls. Strict separation of user data. Administrator dashboard for managing accounts, devices, sessions and usage limits. Server health, queue and failure monitoring. Shared organisation dictionary. Approximately five simultaneously active users with variable workloads. Administrator access should not automatically expose private transcript content. LANGUAGES AND PROCESSING MODES Intended languages: English. Somali. Arabic. Swahili. Required modes: Verbatim transcription. Clean dictation: remove fillers and repetitions, correct punctuation and preserve meaning. Translation into the selected output language. Include automatic language detection and manual language selection. Test accuracy separately for each language and mixed-language speech. Preserve names, numbers, dates, negations and specialist terminology. Cleanup must not invent information or silently summarise speech. AI MODEL RECOMMENDATIONS Please propose the best speech-recognition model and local LLM for this workload. Propose a separate translation model if that would improve results. Explain: Why your recommended models suit these tasks and languages. Expected accuracy, latency and hardware requirements. How they handle accents, background noise and specialist terminology. Whether they support incremental transcription. Their licensing and commercial-use conditions. Trade-offs compared with alternatives. How you will benchmark them on our hardware. Keep model backends replaceable so models can be upgraded without rebuilding the client applications. PERFORMANCE Prioritise interactive dictation over long recordings and background processing. Include fair scheduling, bounded queues, cancellation and timeouts. Prevent duplicate text insertion after retries or connection interruptions. Benchmark one, three and five concurrent users while existing Proxmox workloads are running. Report time to first text, time to final cleaned text, queue delays, accuracy and resource use. Please distinguish measured results from estimates. CONNECTIVITY AND PRIVACY Secure local-network and remote access. Encrypted connections and authenticated users/devices. No external AI APIs or content-bearing analytics. No raw-audio retention by default. Configurable transcript retention and deletion. Clear microphone recording indicators. Disclosure of any external infrastructure dependencies or recurring costs. DELIVERY AND HANDOVER Propose milestone-based delivery, beginning with a working proof of concept. Provide: Windows and Mac applications and an Android installation package. Private backend and administrator dashboard. Full source code and configuration files. Reproducible build and deployment instructions. Backup, update and recovery procedures. User and administrator documentation. Benchmark results, licences and known limitations. PLEASE INCLUDE IN YOUR PROPOSAL Relevant examples of previous work. Your proposed architecture, models and software stack. Experience with Android keyboards, desktop text insertion and self-hosted AI. Exact deliverables, timeline and fixed-price quote by milestone. Hardware requirements and additional or recurring costs. Main technical risks and limitations. Post-launch support options. What you need from us to begin. Our priority is smooth, reliable dictation across devices, with all AI processing hosted on our own hardware.