Eight domains. Every answer a serious enterprise team needs before deploying AI for knowledge transfer at scale.
This document exists for enterprise procurement, legal, and technical teams conducting formal due diligence on Expert Scale. It answers the 30 questions most commonly raised by organizations evaluating AI for knowledge transfer, workforce continuity, and SME scaling.
Every section is written to be shared with internal stakeholders. Nothing here requires an NDA. For financial detail, security architecture deep-dives, or formal questionnaire responses, contact Drew Harris or Alan Kadish directly.
Knowledge preservation for retiring field SMEs is the highest-leverage entry point. When a senior domain expert leaves, the knowledge they carry — the judgment, the heuristics, the pattern recognition that took decades to develop — goes with them. Documentation doesn't capture it. Formal training programs take months and rarely replicate the tacit reasoning. A Digital Protege captures that before the person is gone.
The second-highest-leverage use case is new-hire acceleration. E-Tech Heat Recovery compressed a standard six-month technical ramp to under one month by deploying a senior engineer's Protege as the backbone of onboarding. New hires ask it questions in plain language instead of competing for the engineer's calendar time.
Third: customer-facing expert access at scale. Sales teams and support organizations that need domain expertise in the field — mid-demo, mid-call, mid-proposal — without adding headcount.
Pilot recommendation: start with one retiring or high-demand SME, one well-defined use case (onboarding or field support), 90-day pilot. The ROI data from that deployment sets the business case for scale.
Tacit knowledge capture is the core of the platform. Explicit knowledge — documents, manuals, procedures — is table stakes. Any RAG system can surface a document. The hard problem — and the genuine differentiator — is extracting the "why behind the what": the judgment a precision measurement specialist applies when assessing an ambiguous field condition, the pattern recognition a senior engineer uses to diagnose an edge case before running a formal analysis.
Our extraction methodology draws on 30 years of instructional design research. It is not an open-ended interview. It is a structured AI-driven session designed to surface the reasoning and mental models that domain experts cannot articulate unprompted because they are so deeply internalized. This methodology is patent-pending.
The E-Tech Heat Recovery deployment is the clearest proof of production readiness: a senior engineer with decades of domain-specific hands-on expertise, successfully captured, validated, and deployed for new-hire onboarding. The Protege handles the kind of situational, judgment-dependent questions that no manual answers.
Build — Six-phase process:
Maintain: Weekly expert review requiring approximately 10 minutes. Expert provides natural-language feedback ("say this differently," "always mention X when asked about Y"). Gemini extracts discrete instructions from that feedback, applies them automatically, and regenerates the Protege. No technical configuration required from the expert.
Evolve: Logarithmic improvement curve. Each session compounds depth. Version control on every update. Instant rollback. Nothing goes live without explicit expert approval.
Core differentiator: every answer the Protege gives is grounded in retrieved content from the expert's knowledge base. The LLM is not filling in gaps with general training data. If the knowledge isn't in the KB, the system escalates instead of generating.
Foundation models: Google Gemini (primary, via Vertex AI), Anthropic Claude (advanced reasoning and complex conversation), a voice-AI platform (voice deployment infrastructure), pgvector on Supabase (semantic vector embeddings and retrieval).
Platform is model-agnostic by design. We are not locked to any single model provider. As foundation models improve, Protege capability improves with no rebuild required.
Infrastructure: 25+ microservices on GCP Cloud Functions, coordinated via GCP API Gateway. Each service scales independently. No monolithic bottleneck.
Proprietary layers:
Patent status: 34 claims filed. Independent IP attorney assessment: 17-month minimum engineering effort to replicate the core escalation architecture.
Expert Scale owns and operates:
Customer provides:
Zero customer infrastructure required in the standard SaaS deployment. For organizations with data sovereignty requirements, see Q14.
Core product. Single-expert deployment. Available 24/7 via voice or text. Zero-hallucination architecture. Starts at 50% fidelity after one hour, progresses to 80-90% with enrichment.
Enterprise deployment extension. Designed for organizations of 1,000+ employees. Scales a single expert's knowledge across an entire division. Priced $150K–$500K depending on scope.
Intelligence engine. Powers session analysis, quality scoring, and insight generation across all products. Feeds improvement loops for every Protege in the platform.
Document intelligence layer. Upload standards, bulletins, technical manuals. Protege reasons over them in real time alongside captured expert knowledge. Not just retrieval — contextual reasoning.
Multi-agent orchestration. Coordinates queries across multiple Proteges for cross-team workflows. One question, multiple expert viewpoints, synthesized response with attribution.
Fidelity is defined as the proportion of an expert's domain that the Protege can answer accurately, in the expert's voice, without escalation. 50% means that after one structured one-hour interview, approximately half of the questions a typical user would bring to that expert will receive a grounded, expert-aligned response.
This figure is conservative by design. It is derived from surveys of live experts who assessed their own Proteges — the subject-matter expert is the most demanding possible evaluator. The expert validates their own Protege before it goes live.
The three-stage progression:
After one 1-hour interview. Working Protege deployed same day. Covers core expert judgment and common question types.
After ingesting supplemental materials (manuals, recordings, past work). Fewer escalations, broader domain coverage.
Ongoing sessions, weekly review, feedback cycles. Near-complete domain coverage. Production-grade for regulated contexts.
Pilot integration: we recommend embedding this survey-based validation in pilot success metrics at 30, 60, and 90-day milestones so there is a quantitative anchor for stakeholder reporting.
Improvement is hybrid: AI-assisted session analysis plus active expert curation. Nothing is ingested automatically — this is a deliberate design choice for regulated environments where unvetted knowledge could create liability.
After each session: AI-generated session insights surface what was discussed, what escalated, and what gaps emerged. The expert reviews these and decides what to add to the knowledge base.
Two improvement paths:
Escalation loop: unanswered questions (escalations) are logged with full conversation context. The expert is notified. Their response is fed back into the KB. Each escalation closes a gap.
The result is a logarithmic improvement curve: rapid improvement in early weeks, then continuous refinement as the edge cases get resolved.
Three mechanisms, each independent and usable in combination:
For regulated domains: the right configuration depends on your compliance posture. Two options: (1) updates are silenced until a compliance officer approves before going live, or (2) graceful escalation — the Protege acknowledges the area is under review and routes to a human. This is a configuration choice, not a limitation. We recommend discussing which mode fits your compliance workflow during pilot scoping.
Each Protege is a single-expert system by design. There is no cross-contamination between experts by default. A precision measurement specialist in one division and a calibration engineer in another each have their own fully isolated knowledge environment. Their answers may diverge — that is correct behavior, because their expertise and domain perspective genuinely differ.
In a multi-team enterprise context, this is a feature, not a gap. Each team's Protege carries that team's specific domain knowledge, reflecting how that team actually operates. A synthesized consensus across experts would wash out the very specificity that makes expert knowledge valuable.
When cross-team synthesis is needed: Conductor (multi-agent orchestration, targeted Beta Q3 2026) will allow a single query to route across multiple Proteges, synthesize their responses, and return an attributed answer that shows which expert viewpoint each element came from. This preserves individual expert accountability while enabling cross-domain reasoning.
Until Conductor is available, the recommended workflow is: route to the appropriate Protege for the domain in question. Routing logic can be configured in the platform or handled at the application layer.
Accuracy: retrieval-augmented generation (RAG). Every response the Protege gives is grounded in a retrieved chunk from the expert's knowledge base. The LLM is not drawing on general training data. It is reasoning over content the expert has validated. If the knowledge isn't there, the system does not generate an answer.
Hallucination: the system knows when it doesn't know. Retrieval failure triggers escalation, not a confidence-scored generated response. The Protege routes the unanswered question to the human expert with the full conversation context attached. It does not fabricate. This is the architecture, not a setting.
Liability: Expert Scale does not warrant the accuracy of Protege responses — the expert does, same as for their direct advice. The Protege is an extension of the expert's judgment, not a replacement for it. The expert remains legally responsible for the knowledge they put into the system and validate.
Audit trail: full session transcripts are retained. Two AI-generated summaries are produced after each session — one for the expert's review, one for the client record — timestamped and verifiable. This creates an accountability chain comparable to documentation of direct expert consultation.
The decision to answer or escalate is made at the retrieval layer, not the generation layer. This is the key distinction from systems that generate an answer and then score its confidence.
Flow:
Four architectural guardrails:
Protege asks clarifying questions before answering, ensuring it is reasoning over the right framing of the question.
Session context is maintained throughout the conversation. The Protege does not lose track of prior exchanges and avoids contradicting itself within a session.
The knowledge base is always queried before any LLM reasoning. No general training data fill-in is permitted.
The Protege cannot recommend actions or take positions the expert would not take. Hard behavioral boundaries configured per deployment.
Patent-covered: the escalation mechanism that distinguishes genuine retrieval failure from tangentially related content — preventing false positives (escalating when it should answer) and false negatives (answering when it should escalate) — is covered in the 34 patent claims.
Fully segregated at every layer:
protege_id. Every kbSearch query is scoped at the query level — not just at the API level. There is no code path that can return content from another Protege's namespace.GCP infrastructure certifications (inherited): SOC 2 Type II, ISO 27001, FedRAMP Moderate. Supabase: SOC 2 Type II. These are infrastructure-layer certifications. Expert Scale application-layer SOC 2 is targeted end of 2026 (see Q17).
| Deployment Model | Description | Status |
|---|---|---|
| Standard SaaS | Fully managed on GCP. Logical isolation via RLS and scoped vector search. No customer infrastructure required. Default deployment. | Available now |
| GCP Private VPC | Expert Scale services deployed in a private VPC — either Expert Scale's GCP org or the customer's own GCP project via secure peering. Full network isolation. | By arrangement |
| Customer Infrastructure | Knowledge corpus stored in customer's own infrastructure. Expert Scale connects via encrypted API with dedicated key. Customer retains full physical data control. | By arrangement |
All non-standard deployment models are priced per-engagement and defined in the commercial agreement. Discuss requirements during pilot scoping.
Customer owns all content. All transcripts, uploaded documents, recordings, and derived knowledge (summaries, session logs) belong to the customer organization. This is stated in the contract and is not negotiable.
On embeddings: embeddings are generated using standard Gemini embedding models — not custom-trained models specific to a customer. There is no lock-in at the embedding layer.
Expert Scale retains: the extraction methodology, system prompt architecture, rule engine, escalation logic, and platform source code. This is the intellectual property Expert Scale contributes. The customer is licensing access to these capabilities, not acquiring them.
At termination: customer data is returned in portable format and deleted from Expert Scale infrastructure within the timeframe specified in the commercial agreement. GDPR-compliant deletion procedures. No retention, no negotiation, no lock-in. Specific formats and timelines are defined at the contract stage.
Knowledge leakage between teams is architecturally prevented, not policy-prevented. The distinction matters: policy-based controls can be misconfigured or circumvented. Architecture-based controls cannot.
protege_id. The kbSearch function is hard-coded to search only that Protege's namespace. There is no API path — exposed or internal — that permits a cross-Protege query.There is no shared knowledge pool by default. If a customer explicitly wants to create a shared pool (e.g., a common standards library shared across all teams), that is a deliberate configuration — not a default. Each individual Protege's knowledge remains isolated regardless.
| Certification | Layer | Status |
|---|---|---|
| SOC 2 Type II | GCP infrastructure (inherited) | Certified |
| ISO 27001 | GCP infrastructure (inherited) | Certified |
| SOC 2 Type II | Supabase (inherited) | Certified |
| SOC 2 Type II | Expert Scale application | Targeted end of 2026 |
| ISO 27001 | Expert Scale application | Not currently certified |
| HIPAA BAA | PHI-handling deployments | Available at $25K/yr |
Application-layer SOC 2 Type II is committable as a contractual milestone for enterprise agreements. Both founders have direct experience working in HIPAA-regulated environments. Security questionnaire and architecture review are available for enterprise pilots on request.
Per-SME, grouped by team and business unit. One senior SME per domain per team equals one Protege. A single division might deploy three to five Proteges representing different domain experts (e.g., process engineering, safety compliance, field operations, technical sales). Each team has its own isolated knowledge environment with no knowledge crossing team boundaries.
Recommended sequence:
Multi-expert orchestration (cross-team query coordination) is targeted for Conductor Beta, Q3 2026. Until then, routing to the appropriate Protege for the relevant domain is the recommended approach.
English-first currently. Interviews, knowledge bases, and conversations are conducted in English. This reflects the language of our initial market, not a technical ceiling.
Foundation models (Gemini, Claude) natively support 100+ languages. The voice platform supports multiple languages. Multi-language capability is architecturally achievable with focused development effort.
Roadmap:
If a specific language requirement is material to your deployment, flag it during pilot scoping. We can accelerate roadmap prioritization for enterprise commitments requiring specific language coverage.
Reasons within them, not just references them. The distinction is meaningful. A retrieval system that can surface "ISO 17025 section 5.4.6" is not the same as a Protege that has internalized an expert's standards-informed judgment and can apply it to a novel situation.
The Protege's domain is defined by the expert's knowledge base — which includes everything the expert has internalized about the standards they work within. If an expert reasons in terms of ISO 17025 in every answer they give about measurement uncertainty, that standards-informed judgment is captured in the extraction interview and reproduced in the Protege's responses.
G4 (Scope Fence) enforces this in the other direction: the Protege cannot recommend actions the expert would not recommend. Standards are embedded from the inside (through captured expert judgment), not constrained from the outside.
For literal procedural rules: ingest the standards document as high-weighted reference material in Codex. The Rule Engine supports explicit topic-triggered behavioral constraints (e.g., "whenever the topic is calibration interval selection, always reference the current version of the applicable standard and recommend expert review for non-standard intervals").
Currently founder-led. Drew Harris (CEO) and Dr. Alan Kadish (CSO) handle discovery, scoping, and implementation directly. Enterprise customer success is the first seed-round hiring priority.
Standard implementation timeline:
Average contract-to-live: 3.5 weeks. Best case for non-regulated, well-prepared expert: under 24 hours (Cisco deployed a working Protege in under 24 hours from first interview).
Usage-based licensing at $0.66 per minute ($39.60 per hour). No per-seat fees. No per-Protege license fees. No setup fees in standard pilots. Volume discounts above 200 hours per month.
| Scenario | Sessions/Month | Avg Duration | Est. Monthly |
|---|---|---|---|
| Single Protege pilot | 100 | 20 min | ~$1,320 |
| Single team, 5 Proteges | 500 | 25 min | ~$8,250 + vol. discount |
| 3 divisions, 15 Proteges | 1,500 | 25 min | ~$24,750 + vol. discount |
| Full enterprise | Negotiated | Negotiated | Negotiated rate |
ROI frame: a Protege hour at $39.60 versus a domain expert's billable rate of $200–$500 per hour. The Protege delivers that expert's judgment at 10–20 cents on the dollar, 24/7, with no scheduling lag and full session documentation.
Recommended scope: one team, one high-demand or pre-retirement SME, one primary use case. 90 days.
Measurable pilot metrics:
Non-regulated domain: 3.5 weeks average contract to live. Under 24 hours in best case (Cisco).
Regulated domain: 3–4 weeks, reflecting additional validation steps rather than additional interview volume.
The additional time in regulated deployments accounts for:
The interview itself is still approximately one hour. The additional time is in the validation and staged rollout phases — which are prudent regardless of platform, and which Expert Scale's version control and rollback infrastructure supports directly.
We will give a direct answer to this question, because it is the right question to ask:
Load testing result: 100,000 simultaneous sessions with no infrastructure changes.
To put that in context: a 20,000-person organization where every employee is simultaneously active would generate 20,000 concurrent sessions — one-fifth of the tested capacity. Realistic peak concurrency (5–15% of employees active simultaneously) means 1,000–3,000 concurrent sessions, leaving 97% of tested headroom unused.
Why the architecture holds at scale:
Currently raising $500K seed round via SAFE note at $5.5M post-money valuation. Founder-funded to date.
Financial stability is a fair consideration for any long-term partnership, and we take the question seriously. We are happy to provide detailed financial information, projections, and investor materials under NDA as part of formal due diligence. Contact Drew Harris directly.
Enterprise pilot partners are typically given priority in the investor narrative — and in some cases, enterprise commitments have been structured to directly support runway. We are open to discussing commercial structures that align mutual interests for risk and upside.
Enterprise pilot partners directly shape roadmap prioritization. If a specific capability is material to your deployment, say so during scoping. We build for the use cases that matter.
Foundation model improvements are a tailwind, not a threat. Better base models make every Protege more capable with no rebuild required. Our moat does not depend on the models — it depends on what sits around and above them.
The five layers of durable advantage:
Three operational stages, all fully functional today. Each stage is a complete, deployable state — not a preview.
One structured 1-hour interview. Working Protege deployed same day. Covers core expert judgment, common question types, and the expert's primary domain. Ready for real users.
Add supplemental materials: technical manuals, past project documentation, recorded presentations, regulatory references. Expands domain coverage, reduces escalation rate, increases answer specificity.
Ongoing sessions, weekly expert review, feedback and refinement cycles, escalation loop closing gaps. Near-complete domain coverage. Production-grade for regulated contexts including compliance-critical question types.
The progression from 50% to 90%+ typically takes 60–90 days in a well-supported deployment. The Protege is already useful and deployed at 50%. The question is not "when is it ready" — it is "how much better does it get."
One SME. One use case. 90 days. Every question in this document gets answered concretely by your own data. The founding team is reachable directly below — most pilot scoping conversations take under 30 minutes.
You can also speak directly to a live Protege right now, no scheduling required, to see the technology before any conversation.