Seven architectural moves that separate a multi-tenant, AI-native platform from a course site with AI bolted on. This is the underlying shape, not a tools-I-use list.
One runtime · multiple verticals · same engines for coaching, roleplay, pods & intelligence
The platform runs multiple verticals today, led by GTM OS (customer acquisition) and the 60-Day Founder startup school, alongside Digital Wellness Academy (clinical mental-health education) and others, on shared infrastructure. Same runtime. Same deployment pipeline. Same core engines for coaching, roleplay, pods, and intelligence.
A new vertical means: swap the corpus · retune the prompts · point a new domain · re-brand the surface. Not: rebuild the engine. Each vertical has been ported in about a week.
The compounding effect: every improvement made for one vertical reaches every other vertical on the runtime. When pod-matching gets sharper for GTM OS cohorts, the other verticals get the same improvement on the next deploy.
A single-vertical competitor rebuilding equivalent capability from scratch is racing a treadmill. The platform ships the next vertical when it wants, in about a week, with every prior capability intact.
Most AI products paste a system prompt in front of a user message. This runtime builds a 10+ signal context block on every call: DISC profile, saved artifacts, assessment baseline with scoring reasoning, weak-dimension framework mapping, learning style, current lesson, 7-day outreach stats, full pipeline state, and live semantic search across the user's uploaded documents.
That context is PII-stripped at the boundary by a single centralized function, so every downstream AI call gets the safe version. It isn't per-feature plumbing that can drift, it's one path, shared.
The user-facing consequence: the coach references your deals by name, flags stale pipeline, adapts its tone to your DISC type, and suggests the specific lesson that addresses a specific weak-dimension gap. It's yours, not generic.
The platform isn't a monolith of copy-paste. It's a library of templates, interactive lesson components, prompt templates for coaching personas, facilitator-rhythm prompts, integration adapters, AI flow patterns. Every surface that uses a template inherits edits made to it.
Analytics surface a signal → founder + AI review → a template-level fix lands in one place → the next deploy propagates it everywhere. The human review gate stays in the loop, so AI-driven content changes are supervised, not automated. Competitors on quarterly release cycles can't match this cadence.
That's why "the platform gets smarter while you use it" isn't marketing, it's the operating model. A weakness surfaces Tuesday. The fix is live Thursday. Every lesson, every chat, every roleplay using that template improves at once.
Every third-party integration, CRM, email-finding, exports, messaging, runs on the founder's own API keys, encrypted at rest and scoped per user. The platform routes calls through your accounts; it never holds your vendor relationships hostage to its own. Platform AI is the exception: it's included. You don't bring AI keys, the coaching, roleplay, analysis, and RAG models ship inside the subscription with no per-message markup.
BYOK for integrations isn't a feature, it's a position. When a platform holds your integration keys, it rents your leverage back to you at a markup. This one doesn't. Your contact data stays in your vendor accounts; whatever you already pay Attio, Pipedrive, Brevo, Notion, Hunter, or WhatsApp Business is all you pay.
The side effect: a founder's total cost, platform subscription plus their own integration bills, stays smaller than the all-inclusive price of a tool that bundles those services and marks them up to hedge against power users.
A 7-rule cascading unlock engine routes each founder through the curriculum based on their onboarding assessment, recommended path (inbound / outbound / hybrid), quick wins, critical gaps, journey-map phases, and completion history.
A founder with strong positioning but weak prospecting goes one direction. A founder with the opposite gap goes another. The platform doesn't assume lesson N+1 follows lesson N for everyone, because it doesn't, in real founder journeys.
Paired with the Quick Win Path, a Readiness Score, then the lessons targeting your lowest-scoring dimension, then a usable output in your first session, new founders see value fast. The full library opens up after the first wins land.
Retrieval is two-corpus reciprocal-rank fusion. The platform corpus holds 4,998 chunks from 502 lessons, embedded as vector(1536) in a pgvector Postgres database; the founder corpus holds your uploaded documents. Both are queried independently and merged with RRF at a damping constant of K=60, so the larger lesson corpus can't starve your smaller uploads. The merged chunks are injected into the coaching context on every interaction.
The analytics layer is the training corpus. Five signal streams, completion patterns, coaching conversations, community discussions, assessment performance, manuscript updates, feed a privacy-first improvement pipeline. Founder IDs are cryptographically hashed; names, emails, and chat transcripts never enter the corpus.
The pipeline is built to produce improvements that serve multiple surfaces at once: coaching chat, in-lesson coach, roleplay personas, support-bot first pass, assessment routing, curriculum grounding. One fix lifts the whole surface area, and fine-tunes are published with documented evaluation metrics, peer-reviewable evidence that the AI is infrastructure, not marketing.
4,998 chunks · 502 lessons · embedded vector(1536) in pgvector
Your uploaded documents, queried independently, never drowned out
A single-vertical AI startup amortizes its training-cycle cost over one product. This platform amortizes over many surfaces across multiple verticals on shared infrastructure, structural leverage most AI-native teams can't match.
Every AI task is routed through OpenRouter to a task-appropriate model. The result is a measured 6.2× per-session cost reduction versus a naive all-frontier-model approach, which is exactly why platform AI can be included in the price rather than surcharged.
Persona-voice quality for the coach and the DISC buyer roleplay.
Assessment scoring, ICP validation, quiz reflection, pipeline insights, outreach analysis.
Website analysis, LinkedIn analysis, and RAG retrieval over large inputs.
The highest-stakes assessment scoring, where nuance changes the result.
Called directly, OpenRouter doesn't proxy audio, for synthesis and transcription.
measured per-session cost reduction vs. all-frontier routing
The platform walked away from enterprise-cloud gatekeeping and rebuilt itself in a week on commodity self-hosted infrastructure. Not a stunt, a discipline. Every external edge is designed to be portable.
Deployments are idempotent. Migrations embed in the container entrypoint, no separate migration tool to invoke. Auto-deploy runs on commit to main. Rate-limiting, webhook signature validation, and safe-context extraction live in centralized functions instead of scattered through the codebase.
A two-VPS deployment is live. A dedicated second host runs Uptime Kuma (9 uptime monitors), GlitchTip error tracking (wired to all 4 production applications), and MinIO object storage, fully independent from the primary host, so ops visibility survives any single-host issue. Postgres backups ship cross-host via Dokploy-native pg_dump across 6 databases on a tiered schedule (DWA every 3 hours, GTM OS every 6, the rest every 12), with a 6/6 restore-verify pass confirmed on deployment. Operating cost stays disciplined enough that the margin math works at founder-comfortable pricing, and the savings flow back into capability.
Multi-tenant runtimes exist. Context-aware AI exists. BYOK exists. Template-based architectures exist. Adaptive curricula exist. RAG and improvement pipelines exist. Self-hosted operations exist.
What doesn't exist yet is all seven composed into a single AI-native runtime that a semi-technical founder can operate at a founder-comfortable cost while out-iterating funded teams.
That's what this architecture was designed to produce, and it's why the platform itself is its own Patient Zero.
Tour the real product surfaces, then start a 7-day free trial, no credit card.
BYOK applies only to the third-party integrations you already pay for, Attio, Pipedrive, Brevo, Notion, Hunter, and WhatsApp Business. You connect your own accounts; the platform routes through your credentials with no markup on top of what you already pay those vendors. It does not apply to AI: platform AI, coaching, roleplay, analysis, and the RAG models, is included in the subscription with no per-message markup.
GTM OS uses two-corpus reciprocal-rank-fusion (RRF) retrieval. The platform corpus contains 4,998 chunks from 502 lessons, embedded as vector(1536) in a pgvector Postgres database. The founder corpus contains the documents you upload. Both corpora are queried independently and merged with RRF using a damping constant K=60, so the larger lesson corpus can't starve the smaller founder uploads. The merged chunks are injected into the coaching context on every AI coach interaction.
HITL (human-in-the-loop) template propagation is how platform-authored content and prompt templates are reviewed and approved by a human operator before they reach the live product. When a chapter, a coaching-persona prompt, or a lesson component is revised, the change flows through a review gate before downstream artifacts, coaching context, RAG indexes, lesson content, are rebuilt. Every surface that uses a template inherits the edit on the next deploy, so AI-driven content updates stay supervised rather than automated.
Each AI task is routed to a task-appropriate model via OpenRouter: coaching and roleplay use Moonshot Kimi K2 for persona-voice quality; reasoning-heavy tasks (assessment scoring, ICP validation, quiz reflection, pipeline insights, outreach analysis) use DeepSeek V3.2; long-context extraction (website and LinkedIn analysis, RAG) uses Gemini 2.5 Flash; production-critical assessment scoring uses Gemini 2.5 Pro; voice synthesis and transcription use OpenAI TTS-1 and Whisper-1 directly, since OpenRouter does not proxy audio. This task-strength routing produces a measured 6.2x per-session cost reduction versus a naive all-frontier-model approach, and it is why platform AI can be included in the price rather than surcharged.
The adaptive unlock system gates advanced lessons, workshops, and features against the founder's current assessment scores. If an ICP-clarity score is below threshold, advanced channel-selection lessons stay locked until the ICP workshop is done. It keeps founders from skipping foundational steps that upstream AI flows depend on, and ensures the coaching context is rich before advanced surfaces activate.
Yes. SoloFrameHub runs multiple verticals, led by GTM OS (customer acquisition) and the 60-Day Founder startup school, alongside Digital Wellness Academy and others, on one shared runtime, one deployment pipeline, and the same core engines for coaching, roleplay, pods, and intelligence. A new vertical swaps the corpus, retunes the prompts, points a new domain, and re-brands the surface, it does not rebuild the engine. Every improvement made for one vertical reaches the others on the next deploy.