Institutional Memory Layer: Building Think9's Portfolio Intelligence Infrastructure
A continuous, privacy-governed decision-capture and knowledge-graph engine that turns tacit founder learnings into reusable capital across 30+ co-built brands.
The Problem & Opportunity
Think9 co-builds 30+ brands from zero-to-one, embedded in every founder conversation, vendor call, and pricing decision. That knowledge lives in people's heads, WhatsApp threads, and one-off meetings — not in the organization. The result: brand #31 re-solves problems brand #4 already solved two years ago. As the portfolio scales, this becomes a compounding tax on the one thing Think9's model depends on most — decision speed.
The opportunity is that Think9's edge is a repeatable playbook, not 30 unrelated businesses. That means every captured learning is reusable capital across the entire portfolio, and its value compounds with every new brand added — an institutional memory layer isn't a generic “AI wiki,” it's a direct multiplier on Think9's actual model.
Compounding Knowledge Tax
- •Learnings trapped in WhatsApp threads & non-recorded calls
- •Brand #31 repeating vendor & pricing mistakes of Brand #4
- •Decision velocity slows down exponentially as portfolio expands
Institutional Memory Multiplier
- •Continuous capture mapped to real stakeholder touchpoints
- •Cross-portfolio indexing makes every insight reusable capital
- •Value compounds with every new brand launched in Think9
Enterprise System Architecture & Workflow
1Data sources & Smart Ingestion Router — scoped, not scraped.
High-signal touchpoints upload directly to staging via presigned URLs. A Smart Ingestion Router inspects file types and routes each asset to its optimal processor before vectorization.
Long-form founder calls processed via Google Gemini 2.5 Pro (Google's 1M-token multimodal model) for joint audio transcription + visual on-screen context (e.g. pitch slides, dashboard clicks).
Vendor proposals and operating agreements processed via Google Cloud Document AI Layout Parser (Google's vision-AI model that converts complex PDF tables & contracts into structured Markdown/HTML).
Shared channels where active vendor negotiations and commercial pricing decisions occur.
Procurement pricing tables and SKU metrics inserted into relational tables with automatic row-level vector embeddings.
2Structured Extraction.
Every source passes through a standardized extraction pipeline, producing a canonical record. Entities identified: Brand, Person, Product, Vendor, Consumer Segment, Category. Memory types: Fact, Decision, Insight, Reasoning, Experiment, Outcome, Learning, Action, Risk.
{
"type": "decision",
"title": "Switch packaging vendor for Brand X",
"brand": "Brand X",
"vendor": "Vendor A → Vendor B",
"people": ["Rahul", "Aditi"],
"reasoning": "Vendor A missed 2 launch deadlines in a row",
"evidence": "Meeting #123 (Gemini 2.5 Pro visual chunk), PDF Spec #45 (Layout Parsed)",
"confidence": 0.9
}3Triple-Engine Storage ("Dual Brain" + Graph).
Rather than forcing all content into a generic vector database, we separate structured business data, unstructured document/video intelligence, and decision relationship graphs.
Managed vector store for PDF contracts, meeting transcripts, slide notes, and video visual chapters.
Row-level procurement data, pricing tables, SKU costs, and exact financial metrics.
Multi-hop links connecting Brand → Decision → Experiment → Outcome → Learning.
4Pre-Flight Security Sentinel & Redis Semantic Cache.
Before executing expensive database queries or model calls, incoming questions pass through a Pre-Flight AI Sentinel and a 2-Tier Redis Semantic Cache.
Inspects user input for prompt injection attempts, system overrides, and cross-brand cap-table exfiltration before queries touch database layers.
Stage 1 exact SHA-256 hash match + Stage 2 vector similarity match (>0.85). Returns pre-computed answers instantly for recurring founder questions.
5Federated Router, Neural Re-Ranking & Agentic Toolbelt.
For compound queries (“Have we previously reduced entry pricing for Gen-Z products, and what happened?”), the agent rewrites follow-up turns into standalone queries, fans out across authorized brand vaults, and re-ranks candidate results using a Neural Cross-Encoder.
6Human Governance & Vault-Level RBAC.
Human approval enforces intent fidelity, while explicit Vault Access Policies isolate brand cap tables and commercial terms:
Explicit Vault Access Policies (Read)
Brand-scoped by default via SQL policy filters. Cross-portfolio queries redact sensitive commercial terms to de-identified pattern insights.
Human Approval Queue (Write)
Extracted learnings must be reviewed (approve/edit/reject) by the stakeholder before becoming canonical, preventing memory pollution.
Proof of Concept / Prototype Implementation
To validate the Think9 Institutional Memory & Decision Velocity Engine, we developed a fully functional, end-to-end Proof of Concept (POC) focusing on the highest-leverage operational bottleneck: Cross-Portfolio Procurement & Vendor Intelligence.
The prototype demonstrates how unstructured vendor documents, RFPs, and negotiation touchpoints are transformed into structured database records, governed by human oversight, and indexed for instant, grounded cross-brand retrieval.
Key Architectural Modules Demonstrated in the POC
Multimodal Document Ingestion & Async Processing
- Implementation: Powered by Google’s Gemini API and Node.js Server Actions.
- Workflow: Uploaded Business Requirement Documents (BRDs), Requests for Proposals (RFPs), and vendor quote PDFs are processed via an asynchronous pipeline (
uploadAndWaitForFile). The engine polls file indexing states until active, ensuring heavy document parsing does not block user UI responsiveness.
Automated PO Constraint Extraction & Supplier Compliance Evaluation
- Requirement Extraction: The engine ingests RFP documents and automatically extracts hard operational constraints using
gemini-3.1-flash-lite, grouping them into canonical categories (Product Specifications, Company Qualifications, Delivery & Timelines). - Multi-Supplier Cross-Evaluation: The system parses multiple supplier proposal PDFs against the extracted checklist, generating structured evaluation matrices with explicit status (
MET,NOT_MET,UNCLEAR), exact quotes as evidence, and step-by-step reasoning. - Automated Summary Reports: The system automatically compiles a Markdown Evaluation Report for every PO and syncs it into the RAG vector store.
Dual-Brain Storage (Structured Postgres + Vector RAG Store)
- Relational Memory (Brain B): Stored in Supabase PostgreSQL using normalized tables (
purchase_orders,po_checklists,supplier_submissions,submission_evaluations). Allows exact SQL aggregation for pricing, lead times, and spend analytics. - Unstructured Vector Memory (Brain A): Uses Gemini File Search Stores with
models/gemini-embedding-2for vector similarity and semantic search across supplier PDFs and past negotiation records.
Human-in-the-Loop Governance & Insight Review Queue
- Problem Solved: Prevents “memory pollution” where unverified LLM extractions or raw notes contaminate institutional memory.
- Workflow: When documents are uploaded, the AI automatically extracts key operational learnings (
extractDocumentInsights) and queues them in apendingstate. - Governance UI: Stakeholders review, edit, approve, or reject extracted insights via a dedicated Insights Panel. Only
approvedinsights are promoted to canonical memory.
Grounded RAG Querying with Citation & System Instruction Injection
- Context Injection: During search/chat interactions, all
approvedglobal insights are dynamically injected into the system prompt ofgemini-3.1-flash-lite. - Grounded Answers: The RAG agent executes
file_searchagainst the brand’s store, returning answers grounded in real vendor documents complete with source annotations and citations.
Enterprise Security & Multi-Tenant RBAC
- Authentication: Integrated with Clerk Authentication.
- Vault Scoping: Uses Supabase Row-Level Security (RLS) (
auth.jwt() ->> 'sub' = user_id) to ensure vendor commercial terms, quotes, and cap-table details are strictly isolated per vault/brand.
Technical Stack of the Prototype
| Layer | Technology | Purpose |
|---|---|---|
| Frontend & UI | Next.js 16 (App Router), Tailwind CSS v4, Shadcn UI, Hugeicons | Modern, responsive dashboard with custom PO creation, supplier review modals, and human approval queues. |
| Authentication | Clerk Auth | Enterprise SSO and JWT token issue. |
| Relational DB & RLS | Supabase (PostgreSQL) | Structured data persistence, relational mapping, and Row-Level Security policies. |
| LLM & Vision Engine | Google @google/genai (gemini-3.1-flash-lite) | Checklist extraction, supplier proposal evaluation, insight generation, and RAG chat. |
| Vector Engine | Gemini File Search Stores (gemini-embedding-2) | Managed vector indexing and hybrid search across uploaded PDF assets and reports. |
Demonstrated Impact & Validation
Execution Speed
Vendor proposal evaluation time reduced from 2-3 hours of manual PDF reading to < 15 seconds of automated background extraction.
Zero Context Loss
Learnings from a vendor evaluation in Brand #1 are captured in the human governance queue and made instantly searchable for Brand #30.
Accuracy & Safety
Grounded RAG response generation eliminates hallucinations by linking every recommendation back to verified contract lines and approved insights.
Implementation Plan
Thirty days can't build institutional memory for the full portfolio — the plan is to prove the full loop end-to-end on one function, two touchpoint types, 2–3 pilot brands, with a clear, evidenced path to widen scope after.
Tech Stack Architecture
Unstructured PDF contracts and proposals use Google Cloud Document AI Layout Parser to preserve complex tables without flattening. Long-form founder calls pass to Google Gemini 2.5 Pro for joint audio+visual understanding, cached sub-millisecond via Google Cloud Memorystore, while relational tables in PostgreSQL/AlloyDB maintain exact spend analytics and decision lineage.
Function Scope Decision
Procurement / Vendor Management is the strongest candidate function for the 30 days — vendor decisions are concrete, well-bounded, and directly reusable (a beauty brand's packaging-vendor learning is immediately useful to the next beauty brand entering the portfolio).
Fallback: Marketing / brand knowledge is the fallback if procurement data access proves slower to secure.
30-Day Execution Timeline & Deliverables
Discovery & Scoping
Ingestion & Extraction
Human Review & Intelligence Layer
Test, Measure & Present Rollout
Structured testing on two axes: Accuracy (does the extracted record correctly reflect what was actually decided or said, verified by the pilot stakeholder) and Usefulness (does a query surface something the team didn't already know, or save them re-deriving it).
Fix extraction errors from feedback, and package a phase-2 roadmap for expanding to more departments and brands.
Why a Focused 30-Day MVP Wins
A 30-day MVP that deeply and verifiably nails one function is a stronger proof point than a shallow demo claiming portfolio-wide coverage it can't actually validate in the timeframe — and it directly sets up the case for a longer-term engagement.