The problem
I wanted a voice assistant that could answer finance, real-estate and household questions, but I was not willing to stream my financial life to a cloud provider. The constraint became the design: everything runs on the Spark, and nothing is sent to OpenAI or anyone else.
How it works
Speech comes in through faster-whisper running on the Grace CPU cores, which leaves GPU memory free for the language models. A fast coordinator model handles conversation and decides which specialist should take the request. Specialists answer, and BD speaks the reply back with a synthesised Southern voice.
| Agent | Model tier | Reasoning | What it does |
| BD coordinator | Fast | Off | Routing, persona, quick replies |
| Finance | Heavy | On | Budgeting, statements, tax-year questions |
| Equity analyst | Heavy | On | Runs the quant engine below |
| Real estate | Heavy | On | Rent vs. buy, mortgage and property analysis |
| Auto mechanic | Fast | On | Diagnosis and repair guidance |
| Credentials | Fast | Off | Looks up a password in 1Password after I approve |
Models
| Job | Model | Why |
| Heavy tier | Nemotron 3.5 Super VL 120B-A12B, NVFP4 (75 GB) | Reads charts, 10-K PDFs and screenshots; ships a Spark config |
| Fast tier | Qwen3.6-35B-A3B, NVFP4 | About 92 tokens/s on one Spark; strong tool calling |
| Ears | faster-whisper large-v3-turbo | Runs on CPU so the GPU stays free |
| Voice | Qwen3-TTS 1.7B VoiceDesign | Designs a voice from a written description; no real person is cloned |
Three profiles trade quality against speed: deep runs the heavy model for every agent at 70% of memory, fast runs the small model everywhere, and two-spark splits the tiers across two machines.
The quant engine
- Financial statement analysis: Piotroski F-score, Altman Z, ROIC, DuPont ROE, accruals ratio, free-cash-flow conversion and yield, and automatic red flags.
- Index valuation: a 0–100 overvaluation score for VOO and QQQ built from CAPE percentile (35%), equity risk premium vs. the 10-year (25%), top-holdings richness (15%), trend extension (15%) and concentration (10%). Above 55 it ranks ten alternatives by valuation and trend.
- Stock screen: S&P 500 ranked on momentum (40%), quality (25%), value (20%) and low volatility (15%), dropping names with two or more red flags, with position sizes capped at 12%.
Honest results
A 2010–2026 backtest of the screen (top 20, monthly, regime filter, 10 bps costs) returned a 14.0% CAGR against 14.4% for SPY, with a Sharpe of 0.89 vs. 1.02 and a similar maximum drawdown. Even with survivorship bias in its favour, it did not beat the index. I treat its picks as a research list, not a shortcut, and improving that edge is the next piece of work.
Security choices
- Passwords are copied to the clipboard only after I approve, and never pass through a model.
- The 1Password vault BD can see excludes email, banking and recovery logins.
- No cloud APIs, no telemetry. The machine can be unplugged from the internet and still work.