realdrseattle® AI VenturesFour AI products.
One self-improving machine.
We don't just market with AI —we build it from the metal up. Four GPU-native products,
each running its own intelligence on owned hardware, each protected by hard-coded rules the model
can't override, each retraining itself nightly on verified outcomes.
01
TradingACT.ai
Autonomous Capital Trader.Two locally-trained LLM brains reason over markets 24/7, execute through 7 hard-coded risk gates, and retrain nightly on verified trade outcomes — entirely on owned NVIDIA hardware.
203Klines of Python1,857commits7B + 30Bdual brain34subsystems24/7multi-GPU fleet
Dual brain
A 7B Scanner screens every tick — "is anything brewing?" A 30B-A3B MoE Analyst wakes on signal, reasons across 30+ tools, and emits a structured trade plan: entry, stop, target, thesis.
Risk the AI can't break
7 hard-coded gates stand between the model and capital. Any one vetoes — no override, no exception. Authority · Conviction · Readiness · Champion · Risk · Regime · Calibration.
Zero cloud cost
Scanner + Analyst served locally via Ollama on owned GPUs. No per-token fees, no rate limits — the only viable economics for reasoning every 60–180 seconds, around the clock.
Self-improving loop
Every closed trade becomes a labelled example → reward signal → preference pairs → nightly LoRA retrain → champion gate → hot-swap. The model compounds every 24 hours.
Explore→Plan→Act→Verify→Learn
02
SocialJaaspire
GPU-native social platform.NVENC hardware transcoding and a locally-served Qwen-30B run simultaneously on the same owned GPU stack — zero cloud-LLM dependency, zero third-party token cost.
NVENC / NVDECtranscodeQwen-30Bdual-role7workload gatesPWA + Weblive users
Dual engine
NVENC/NVDEC delivers multi-bitrate streams the instant a creator uploads — no CPU farm, no cloud transcode bill. One Qwen-30B plays two roles on the same GPU: Guard and Companion.
Guard-first publish
No post reaches the public feed until Qwen-30B moderation clears it. Safety is enforced before publish, not patched after.
Monetization the AI can't bypass
Wallet-gated previews, per-creator post tiers (regular / paid / subscription), and 2FA — all server-enforced, all on owned infrastructure with zero third-party egress.
The flywheel
Every moderation verdict and chat session becomes labelled data → nightly LoRA fine-tune of Qwen-30B → champion-gated hot-swap. Real creator behavior, not synthetic prompts.
Upload→Encode→Guard→Publish→Engage
03
SEODr. SEO
AI SEO audit platform that runs entirely on-premise— citation-enforced, local-LLM, and compounding with every audit. The only architecture for regulated markets that can't ship data to the cloud.
83KLOC · TypeScript104audit modules62-page dashboard~1.5KLLM passes / audit
GEO tracking
Multi-engine citation tracking across Perplexity, ChatGPT, Gemini, Bing Copilot and AI Overviews — the new SEO surface no incumbent measures.
Site memory
A per-domain personality profile is written after each crawl. The next audit reads it back, biasing strategy toward what already worked for that site.
100% on-premise
1,300–1,500 LLM passes plus 5,000+ embeddings per audit — all local. It unblocks legal, healthcare and finance teams who can't send data to OpenAI.
7 grounding gates
Citation · Consensus · Provenance · Fallback · Champion · Concurrency · Hash cache. Any gate fails → the recommendation is scrubbed, not downgraded.
Crawl→Council→7 Gates→Fine-tune
04
AgentsConvoia AI
The Agent Fleet.12+ production agents serving teams across legal, medical, marketing, engineering and customer ops — live on the Play Store today, now migrating each agent to a specialized, GPU-native, self-improving LoRA adapter.
12+production agentsQwen3-14Bshared baseLoRAper agentPlay Store· May 2026
Specialized, not prompted
A legal agent on a generic model is just that model with a system prompt. Each Convoia agent is LoRA-trained on its own domain — a+25–40% accuracy targetover the base.
Privacy opens verticals
Legal, medical and financial customers can't ship sensitive prompts to cloud APIs. On-premise agents open entire regulated verticals that were previously off-limits.
Cost that scales
High-volume agents are loss-making at per-token cloud rates. Bringing them home drops cost to amortized GPU — and targets<200ms TTFTvs an 800ms–3s cloud round-trip.
Compounds nightly
Customer corrections and expert annotations feed a nightly LoRA retrain → champion gate → adapter hot-swap. Every interaction sharpens the next.
Ingest→Specialize→Serve→Verify→Improve