KnowledgeVault AI research
internal prototype · canonical JSON + Dreamborn Forge HTML
internal generated
research · supabase_json

KnowledgeVault AI research

research artifact · for KnowledgeVault AI · status approved

Planning Surface

Use this to decide what happens next.

Status

approved

Agent Handoff
Start Here

research artifact · for KnowledgeVault AI · status approved

Completion Evidence

No explicit evidence field yet. Require tests, screenshots, linked PRs, or reviewed outputs before marking complete.

Artifact Shape
  • project: KnowledgeVault AI
  • version: 1.1
  • demo flow: steps: object, format: string, the_gut_punch: string, total_duration: string, target_audience: string
  • generated: 2026-04-29
  • key stats: 5 items
  • pages needed: how_it_works: string, selling_page: string, expert_signup: string, marketplace_home: string, company_dashboard: string, capture_experience: string
  • user stories: the_buyer_story: object, the_plumber_demo: object
  • market context: the_crisis: string, market_size: string, why_distribution: string
  • the dual narrative: the_ask: string, real_story: string, surface_story: string, what_ad_cares_about: object
  • competitive landscape: the_gap: string, existing_tools: object, the_fundamental_flaw: string, why_they_fail_in_distribution: string
  • architecture decisions: locked: boolean, review_sources: object, virtual_expert: object, payment_integrity: object, model_architecture: object, session_context_object: object, structured_decomposition: object
  • independent ai reviews: openai_o3: object, gemini_2_5_pro: object
Structured Payload

Machine-readable source fields

project

KnowledgeVault AI

version

1.1

demo flow
steps
stepactoractionduration s
1expertOpens app, scans Wilkins 600XL PRV barcode, taps START. Meter shows surge badge.10
2systemGPT-4o loads Gemini pre-seed. First question fires — deeply specific to this SKU variant. Expert says: "In 40 years, no apprentice has ever asked me that."20
3expertAnswers by voice. AI detects hesitation on a field tip. Follow-up fires. Expert describes the hum that tells you it is working.30
4systemLeft panel fills in real-time: failure_mode, psi_range, field_tip extracted as micro-objects as he speaks. Meter ticks: +$3.20 earned.10
5systemSession closes. Validation badge: Accuracy 94% — verified against Zurn spec sheet. Rarity bonus fires: novel term field_tip:hum verified and not in base knowledge set.5
6field_techNew hire on job site types: "Why is 600XL chattering at 40 psi?" — the just-captured answer appears at top, tagged field-verified 3 minutes ago.15
7systemAnalytics panel: 119 hours of retiree knowledge preserved this month. -18% predicted reduction in tech-support calls. ROI gauge green.10
format

one uninterrupted flow — capture, structuring, instant retrieval

the gut punch

The field tech finds the answer the expert just gave, thirty seconds later. That is the moment for a room of distribution executives.

total duration

under 3 minutes

target audience

1,500 AD distribution executives

generated

2026-04-29

key stats
  • 41% of organizations never collect knowledge from retiring employees (APQC)
  • 10,000 boomers retire daily until 2030
  • 2.8 million distribution and manufacturing jobs lost to retirement by 2033
  • AD members: $100B in annual sales, 1,500 distributors, 13 industries
  • Gig economy: $2.7T market, fastest growth in specialized knowledge work
pages needed
how it works

The AI-first flow explained. Expert side and buyer side. The quality economy.

selling page

For AD / buying group decision makers. The platform story. The build speed and cost story. The partnership ask. This page sells BezelIQ as much as KnowledgeVault.

expert signup

Tradesman onboarding. Role selection, terms, rate display, bank stub. Mobile-first.

marketplace home

What KnowledgeVault is, who it is for, why it is different. For both experts and buyers. Clean, Blanc-based, logo-swappable.

company dashboard

Static coverage report. No auth. Loads in under 2 seconds on a projected screen.

capture experience

The hero demo. Voice-first AI interview. Mobile, full-screen, one question at a time.

user stories
the buyer story
sees
  • Coverage grid: 88 SKUs, green/yellow/red by capture status
  • Slang cloud: the actual terms field techs use
  • Top searched topics across captured knowledge
  • Virtual Expert chatbot trained on captured knowledge — answers field questions with real experience, not spec sheets
scene

Operations manager, did not participate in capture sessions. Just bought access.

the plumber demo
flow
  • Opens app. Sees queue: 3 uncaptured products in his specialty.
  • Taps a Watts pressure reducing valve. Sees product photo, SKU, surge badge.
  • Taps START. GPT-4o loads Gemini pre-seed for this SKU. First question fires — deeply specific to this variant.
  • Answers by voice. Transcript appears live. He never types.
  • AI hears hesitation and asks a follow-up: "You mentioned it needs to be set before commissioning — what happens if a contractor skips that step?"
  • He answers. That answer captures the failure mode new techs get wrong every time.
  • After 6 base questions + 2 AI follow-ups, hits DONE.
  • Session complete: earnings ($16.80 after quality multiplier), scores, badges. Rarity bonus fires — "chattering valve" verified against spec, not in base knowledge set.
  • 8 structured knowledge micro-objects created in 20 minutes.
scene

A retired plumber, 62 years old, on his couch. He pulls up KnowledgeVault on his phone. He made $80 last week in two hours.

what makes it ai first

A form would have gotten 6 generic answers. The AI heard a signal and dug deeper. That follow-up is the difference between useful knowledge and a completed form. No human interviewer could do this at scale for every SKU in every distributor catalog.

market context
the crisis

The manufacturing and distribution industry is in the middle of a knowledge cliff. 4 million Americans turn 65 every year through 2027. 10,000 boomers retire daily until 2030. Manufacturing alone loses 11.8% of its workforce — 1.841 million workers. 2.8 million of 3.8 million total job openings through 2033 are retirement-driven. The killer stat: 41% of organizations rarely or never collect knowledge from retiring employees. That knowledge is not in any document. It lives only in someone's head.

market size

Affiliated Distributors: 1,500+ independent member distributors, $100B in 2025 member sales (record), 13 divisions, 3 countries. AD is actively pushing digital transformation tools to members — this is a warm market with a named contact (Brian Radichel).

why distribution

Distribution is hardest hit. A plumbing distributor's best asset is a 30-year field tech who knows every valve in their catalog — which connections blow out in cold climates, what tradespeople actually call the fittings on a job site, what the failure modes look like before they fail. None of that is in a spec sheet. When that person retires, it walks out the door.

the dual narrative
the ask

Not: buy KnowledgeVault for $X/month. The ask is: partner with BezelIQ to bring AI-first tools to your members. KnowledgeVault is the first product. Pricing intelligence, CSR augmentation, quote automation are next. AD is the distribution channel. BezelIQ builds the tools.

real story

The platform that built this in 6 days is what Justin is actually selling. The demo is not a product demo — it is a capability demo. At the end, the conversation shifts: "This took us 6 days. Here is what it cost us to build. Here is what a dev shop would have charged. What do you want us to build next for your members?"

surface story

KnowledgeVault: capture expert knowledge before it retires. AD sells to 1,500 member distributors. Every member has retiring experts.

what ad cares about
  • Their 1,500 members all face the same knowledge cliff
  • AD's value is giving independent members tools national chains do not have
  • A platform that builds custom AI tools for any member use case at a fraction of enterprise cost is AD's competitive moat for the next decade
competitive landscape
the gap

There is no product on the market that extracts tacit expert knowledge through AI-driven conversation and converts it into structured, searchable assets at scale. That gap is KnowledgeVault.

existing tools
  • Guru
  • Bloomfire
  • Tettra
  • Confluence
  • Notion
the fundamental flaw

Every tool in this market has the same flaw: they require the expert to know what to document. A wiki assumes the expert can write. A FAQ system assumes someone already knows which questions to ask. A LMS assumes the knowledge already exists in structured form. None of them extract tacit knowledge — the stuff experts know but cannot articulate until asked exactly the right question.

why they fail in distribution

A retiring plumber cannot write a knowledge base article. He has never done it, does not have time, and does not know which of the 10,000 things in his head are valuable. Guru and Bloomfire are tools for knowledge workers. Distribution experts are tradespeople.

architecture decisions
locked

true

review sources
  • Gemini 2.5 Pro (2026-04-29)
  • OpenAI o3 (2026-04-29)
virtual expert
decision

Expert is a first-class product entity, not a tag on a knowledge chunk.

description

A Virtual Expert is a named, queryable AI persona derived from capture sessions with a real person. Field technicians query the Virtual Expert directly. The retiring expert never really leaves.

implication

Schema must model Expert, CaptureSession, and KnowledgeChunk with expert_id FK. The product is sold as Virtual Expert access, not knowledge base subscriptions.

payment integrity
gate

o3 cross-checks novel terms and claims against manufacturer manuals, plumbing codes, and distributor ERP master data before applying rarity or quality bonus.

decision

Ground-truth gate is mandatory before any payment_modifier > 1.0 fires.

rationale

Novel terms earn more without verification incentivizes jargon invention. Experts must feel the scoring is fair and ungameable or they leave.

failure mode

Novel terms that fail spec cross-check earn base rate only. Claims that contradict ground truth trigger payment hold pending manual review.

model architecture
phases
inputmodelphaseoutputtimingrationale
- product_image - catalog_description - spec_sheet - manufacturer - category - part_numberGemini 2.5 Propre_processing- opening_question_set (3-5 questions ordered by knowledge depth) - key_knowledge_vectors - product_context_summaryoffline, per SKUGemini multimodal depth applied offline eliminates real-time latency. For 80 demo SKUs, pre-run once and cache. Re-run per SKU when specs update.
- pre_seeded_question_set - product_context_summary - live_audio_transcriptGPT-4ointerview- conversation_transcript - follow_up_questions - extracted_micro_objects - session_context_objectreal-timeSingle model in the real-time loop eliminates cognitive dissonance and latency. GPT-4o fast, reliable, multimodal. One fine-tuning target over time.
- full_session_context_object - transcript - extracted_micro_objects - product_spec_for_ground_trutho3evaluation- score_object (7 dimensions) - payment_calculation - ground_truth_validation_resultasync, post-sessionAsync removes latency constraint. o3 deep reasoning appropriate for payment-critical scoring. Never in real-time path.
decision

Three-phase, two-role architecture. Single real-time model. No multi-model chain.

why not claude

Anthropic TOS may prohibit piece-work pay arrangements where Claude drives human compensation. Keep Claude out of the payment path.

session context object
fields
  • session_id
  • expert_id
  • sku
  • product_context_summary
  • question_sequence
  • transcript
  • extracted_micro_objects
  • audio_signals_flagged
  • conversation_arc
decision

A SharedExpertContext object is passed between all three phases. The evaluator scores in context, not blind.

structured decomposition
who

GPT-4o extracts these during the interview in real time. They appear in the demo structuring panel as the expert speaks.

fields
  • sku
  • failure_mode
  • symptom
  • fix
  • psi_range_or_specs
  • safety_warning
  • field_tip
  • tool_required
  • part_number_reference
decision

Captured answers must be decomposed into structured micro-objects in real time, not stored as transcript text only.

rationale

Transcript text is not searchable or reusable. Micro-objects power field queries, counter-sales chatbot, and AR overlays. This is the long-term moat.

independent ai reviews
openai o3
date

2026-04-29

key quote

Turn tacit voice notes into machine-addressable micro-objects in minutes, not weeks. Nobody else in B2B distribution has that.

top findings
  • Novel terms earn more without verification incentivizes jargon invention — ground-truth gate required
  • Answers need structured decomposition into micro-objects (SKU, failure mode, torque spec, safety warning)
  • No human-in-the-loop safety valve — payment and legal liability on one automated score is risky
  • Freeze evaluator model version when payments are live — treat updates like a tax rule change
  • ASR error bias against accents lowers scores unfairly — WER compensation needed
scoring rubric
dimensions
nameweight
technical_accuracy0.3
completeness0.2
practicality_tacit_insight0.15
safety_compliance0.15
clarity_structure0.1
novelty0.05
brevity_efficiency0.05
gemini 2 5 pro
date

2026-04-29

key quote

You are building a system with a permanently fractured brain.

top findings
  • Three-model real-time chain is brittle and high-latency — consolidate or pre-seed
  • Models do not share state — scorer evaluates decontextualized text without knowing what question was asked
  • Virtual Expert is the missing product insight — field techs query Bob directly, not a knowledge base
  • Demo must show capture then the AI as that expert solving a real crisis — two acts
ai first model inversion
pay signals
  • Base rate: $2.00 per answered question
  • Quality multiplier: AI score 0-100, applied to base rate (7-dimension rubric)
  • Rarity bonus: novel terms not in existing knowledge base — requires ground-truth gate before firing
  • Surge bonus: high-demand SKUs with knowledge gaps
  • Completion bonus: first expert to fully capture an uncaptured product
the inversion

Traditional SaaS: company pays for the platform, users fill it with content for free. KnowledgeVault: company pays for captured knowledge assets, experts earn for creating them. The AI is the quality judge that makes this economically viable at scale.

what saas does

SaaS knowledge tools charge a flat subscription. Value is proportional to how much humans manually put in. You pay $40/user/month whether the content is good or garbage. The model rewards volume, not quality.

what ai first does

KnowledgeVault pays experts based on quality as evaluated by AI in real time. The AI scores depth, specificity, and novelty. A plumber who says "shark bite" instead of "push-to-connect fitting" earns more — because that slang is what the next tech will search for on a job site. The model rewards signal, not noise.

why experts show up

Legacy + income. Baby boomers want their expertise recognized and preserved. The platform gives them a professional profile and a record that their knowledge contributed to the industry. The money matters, but the legacy matters more.

open questions for blueprint
  • How is build speed/cost surfaced in the demo — hidden meta-page, live cost counter, or talking point only?
  • What does the AD reseller model look like — white-label, powered-by badge, or private label entirely?
  • Mobile-first for expert capture — offline support scope for demo?
  • Is the selling page a separate domain (e.g. knowledgevaultai.com) or part of the app?
  • Tech stack: Blanc CSS + vanilla JS for marketing pages, Next.js for the capture app — confirm split or unified?
  • Voice cloning for Virtual Expert demo (Gemini suggestion) — in scope for demo or V2?
  • Counter-sales integration (Virtual Expert surfaces reorder recommendation) — demo feature or V2?
  • ASR accent bias compensation — in scope for demo or Policy-level concern?