Built by hand. Deployed to production.
A personal portfolio of AI systems designed, built, and deployed end-to-end by Nicholas Falshaw. Each one runs in production — self-hosted, Dockerised, or packaged as a desktop app. None of this is aspirational. Every entry below is working code.
Published as evidence of independent engineering work — the kind of systems I build outside of any employer's scope.
LLM Fine-Tuning Pipeline →
End-to-end custom model training, delivered as Docker
A production fine-tuning pipeline for small-language models. Client data is prepared locally in Alpaca/ShareGPT format, trained in the cloud via LlamaFactory on GPU, converted to GGUF, then packaged into a Docker image the end user can run entirely on-premise with Ollama.
Highlights
- ▸Data prep then cloud train then GGUF convert then Docker package
- ▸Tracked experiments with MLflow
- ▸Runs on-premise for end user after training
Stack
finetune.rogueai.de/
Pension AI →
AI-powered retirement planning with Monte Carlo simulations
Retirement planning platform combining deterministic portfolio math with an LLM financial advisor. Users upload current holdings, define retirement goals, and receive probabilistic projections plus narrative analysis grounded in their own data via pgvector retrieval.
Highlights
- ▸Monte Carlo portfolio simulations
- ▸Goal-based retirement planning with AI advisor
- ▸100 percent containerized; runs on a single VPS
Stack
pension.rogueai.de/
Tax Document AI →
OCR + LLM scanner for small-business tax filing
Self-hosted tax document manager. Receipts and invoices are scanned with llava:13b for multimodal OCR, then classified and summarised by llama3.2. Designed for Bulgarian SME tax filing; architecture is jurisdiction-agnostic.
Highlights
- ▸Multimodal OCR via local llava:13b
- ▸Categorises receipts, invoices, contracts
- ▸Zero-cloud processing — documents never leave the host
Stack
taxdoc.rogueai.de/
VarnaAI CRM →
Personal CRM with AI-assisted workflows
A contacts / companies / leads / deals / events / tasks CRM built from scratch. Session-scoped auth, Prisma 7 with PrismaPg adapter, and Ollama-powered assistance on every record. Replaces commercial CRMs for independent consultants.
Highlights
- ▸30+ API routes with per-user data scoping
- ▸AI assistance via Ollama on contacts, deals, and event notes
- ▸Playwright E2E test suite with coverage reporting
Stack
crm.rogueai.de/
Meeting Copilot →
Live transcription + Claude whisper in a side panel
A desktop meeting assistant running as an Electron tray app. Audio is captured from both the microphone and system loopback, streamed to Deepgram for transcription with speaker diarization, and piped into the Claude Agent SDK so the user can ask questions about what was just said mid-meeting.
Highlights
- ▸Captures mic + system audio, mixed to 16 kHz mono
- ▸Streaming transcription with diarization
- ▸F10 hotkey queries Claude about the live transcript
Stack
meeting.rogueai.de/
Offline Voice Dictation →
System-wide push-to-talk dictation, zero network
A system tray app that brings offline voice input to every application on Windows. Audio recorded via MediaRecorder at 16 kHz, transcribed locally with a quantised Whisper-small ONNX model, and injected into the active window via clipboard + synthesised Ctrl+V.
Highlights
- ▸100 percent offline — model runs in a renderer Web Worker
- ▸F9 push-to-talk, pastes into any focused window
- ▸~2-4 second latency on CPU inference
Stack
voice.rogueai.de/
20+ additional internal AI systems built end-to-end — from RAG pipelines serving compliance teams to intelligence briefing platforms. Catalogue available on request.