Decision Intelligence Platform
Complete Build Plan
Everything required to go from the current working system on the DGX to a fully licensed, production-grade multi-tenant product — all capabilities, all gaps, all costs, one picture.
Live and operational
Corpus ingestion pipeline
LiveBook ingest → text extraction → chunking → trust classification → embedding. 298 books tracked, 190 trusted and embedded.
Semantic retrieval engine
LiveCosine similarity over 84K chunks. Query-time embedding via qwen2.5:3b. Top-60 expansion, reranked to final-12 source pack.
GDELT live context injection
LiveReal-time geopolitical signals injected per case. Default-on. Non-fatal fallback if GDELT unavailable.
Local option generation
Liveqwen2.5:14b generates strategic options from retrieved context. Runs entirely on DGX — zero API cost.
Sonnet synthesis
LiveAnthropic Sonnet (customer key) synthesises final strategy output with source attribution and recommendations.
Queue-based job processing
LiveCloudflare KV trigger → systemd queue watcher → pipeline → KV result. Browser polls every 8s. Event-driven, no polling loops.
Web interface + CRM
Livevisuals.professionalopinions.com.au. Scenario tester, 5 strategic modes, tension controls, tester enrolment. Attio CRM sync on registration.
Audience projection layer
LiveSegment-level heuristic audience responses. Scaffolded for future ABS/polling integration. Runs per case.
Attribution + compensation ledger
LiveChunk-level influence tracking. Revenue split model: 30% strategist pool, distributed by influence weight × reputation multiplier.
Report rendering
LiveHTML, Markdown, JSON, and frontier report bundle. Export-ready artifacts per case.
Partial / in progress
Master strategist evaluation tier
Premium — ready, no enrollees7-tier architecture designed. 10 profiling scenarios across 3 tiers operational. capture_deliberation.py and aggregate_strategist_profile.py built. Awaiting first strategist enrolment and session capture.
106 quarantined books
PartialContent exists, not yet trusted. Phase 1-2-3 rehabilitation pipeline built. If rehabilitated: estimated +40,000–60,000 additional chunks, nearly doubling retrieval depth.
Live research bundle
Partialbuild_research_bundle.py stores web summaries and notes per case. Not yet automated — web search/fetch done manually. Next step: auto-search + GPT synthesis.
Preference profile feedback loop
PartialUser preference profiles created on enrolment. Selection logging built (log_strategy_interaction.py, update_preference_weights.py). Not yet wired to option generation weighting.
Outcome feedback (30/60/90 day)
Partialoutcome_feedback.py built and schema-ready. Not yet integrated into the web interface or automated follow-up pipeline.
PDF export
Partialrender_capability_manual_pdf.py exists. Not yet integrated for per-case report export or canary-marked outbound briefing.
Not yet built (identified gaps)
ABS + polling data integration
GapAudience projection is heuristic-only. No ABS demographic data ingestion, no polling API connection, no sentiment trend ingestion. Required for quantitative projections.
Multi-tenancy + auth layer
GapNo customer accounts, no isolated data namespaces, no per-customer API key vault. Single-operator system only. This is the primary blocker for licensing.
Token/cost dashboard (customer-facing)
GapNo customer-visible usage or cost tracking. Required for licensed product where customers use their own Anthropic keys.
Strategist web portal
GapNo web interface for premium tier strategists to review their profiles, see attribution, or track compensation. Currently CLI-only.
Voice transcription integration
GapStrategist profiling sessions require transcript file input. Voice-to-text not automated. Limits profiling to text-only sessions.
Multi-domain packaging
GapPlatform architecture supports other decision domains beyond political strategy. No modular packaging system to deploy a second knowledgebase (e.g. insurance, finance, health policy).
Canary-marked briefing exports
GapSteganographic email fingerprinting for leak attribution exists in the workspace but is not integrated into the strategy report export pipeline.
The platform does not need to be complete before generating revenue. Twelve discrete milestones, each independently launchable and independently monetisable. Build in sequence — earlier milestones fund later ones and each one leaves the system better than before.
Corpus rehabilitation
Run the Phase 1-2-3 rehabilitation pipeline on 106 quarantined books. Already scripted. Near-zero effort, high impact — expected +40–60K additional retrieval chunks, nearly doubling depth. Every case run after this is materially better.
Containerise + PDF export + canary reports
Dockerise all services. Integrate per-case PDF rendering. Add steganographic canary marking to all outbound briefings. The system now produces a professional, leak-attributable deliverable that looks like a real product.
Pilot customer — single operator, manual billing
The system is already good enough to charge for. Identify one known person — political advisor, campaign director, consultancy — who has a live strategic problem. Run cases for them directly. Invoice manually. No licensing infrastructure required. This is the fastest possible path to paid validation.
First master strategist — recruit and begin profiling
Start recruiting now. Profiling requires 4+ sessions over 4–6 weeks and cannot be compressed — you cannot start it later and have it ready sooner. This track runs in parallel with all infrastructure work. A single active strategist profile can be sold to the pilot customer as a premium add-on before multi-tenancy exists.
Automated live research + preference feedback loop
Upgrade build_research_bundle.py to auto-search and synthesise current context. Wire preference weight updates into option generation. The platform now learns each user's strategic style over time and removes manual research prep from each case.
Auth layer + per-customer key vault
JWT auth, customer accounts in Postgres, AWS Secrets Manager for Anthropic key storage. This is the minimum viable multi-tenancy — enough to safely take a second and third customer without data bleed, even before the full customer interface is built. Manual onboarding, manual billing still acceptable at this stage.
AWS deployment + customer interface + token dashboard
ECS cluster, RDS Postgres multi-AZ, SQS queue. Customer-facing web portal with case history, usage dashboard, and token/cost visibility. Stripe billing. Self-service onboarding. This is the milestone that converts the system from an operator-managed service into a self-serve licensed product.
Strategist network at scale — 3–5 profiles + portal
Enrol 3–5 strategists across operative, academic, and specialist categories. Build the strategist web portal (view profile, attribution, compensation). Add voice transcription for session capture. Chunk-level attribution for higher compensation accuracy. Three active profiles is the minimum credible premium tier.
Self-healing, hardening, load testing
Circuit breakers on all external APIs. Queue-based async throughout. PagerDuty alerting. Automated DB backup with tested restore. Load test at 10× expected peak. Security audit. This is the milestone that makes the platform genuinely mission-critical grade — required before selling to government or enterprise clients.
ABS demographic integration + polling data layer
Ingest ABS census and demographic data. Connect published polling (Newspoll, Essential, Resolve). Replace heuristic audience projections with population-backed estimates. This is a discrete data engineering module that can ship independently of everything else and be positioned as an "intelligence upgrade" to existing customers.
Modular knowledgebase + second domain
Abstract the political strategy corpus and domain prompts into a pluggable module. Deploy a second domain knowledgebase (e.g. insurance risk strategy, financial services, health policy). Each new domain is a new product line on the same infrastructure, with its own corpus, its own strategist network, and its own pricing.
White-label enterprise deployment
Custom domain, custom interface, proprietary knowledgebase from client's own documents. Dedicated instance. One enterprise client at this tier covers the cost of 20+ base-tier customers. Target: major consultancies, political parties, government departments, financial institutions with proprietary research they want to make queryable.
Build cost and revenue unlock per milestone
Solo costs are near-zero cash (your time). Contractor costs are AUD at $150–250/hr senior DevOps/fullstack. Strategist session fees are additional. Running infrastructure: $165–1,370 USD/month depending on scale (see Deployment page).
| Milestone | When | Solo cash | Contractor (AUD) | Revenue unlocked |
|---|---|---|---|---|
| M1 · Corpus rehab | Wk 1–2 | $0 | $0 | Better outputs — no direct revenue yet |
| M2 · Containerise + PDF + canary | Wk 2–4 | $0 | $7–13K | Professional deliverable; fundable as consulting |
| M3 · First pilot customer | Wk 3–6 | $0 | $0 extra | $500–2,000/mo first revenue |
| M4 · First strategist (parallel) | Wk 2–12 | $2–10K session fees | $3–5K facilitation | Premium add-on to pilot customer |
| M5 · Auto research + feedback loop | Wk 6–10 | $0 | $5–10K | Quality uplift; premium feature positioning |
| M6 · Auth + key vault (min. multi-tenancy) | Wk 8–14 | $0 | $11–19K | Customers 2 + 3: +$1–4K/mo |
| M7 · AWS + customer UI + billing | Wk 12–18 | $200–400 AWS | $23–44K | General licensing: self-serve onboarding |
| M8 · Strategist network (parallel) | Wk 10–20 | $6–30K session fees | $18–31K | Premium tier: $1,500–3,000/mo per customer |
| M9 · Self-healing + hardening | Wk 20–26 | $500–2K audit | $10–18K | Government + enterprise contracts unlocked |
| M10 · ABS + polling data | Mo 6–9 | $500–5K/yr data | $13–24K | Higher price justification; govt contracts |
| M11 · Second domain | Mo 8–12 | $0 | $32–60K | New market: insurance / finance / health |
| M12 · White-label enterprise | Mo 12+ | $0 | $10–18K | $10–30K/mo per enterprise client |
| Full platform (M1–M12) | 24–36 months | ~$9–47K AUD | ~$132–242K AUD | Multi-domain licensed product + enterprise tier |
| To first revenue (M1–M3) | 6 weeks | ~$0 | $7–13K | $500–2,000/mo pilot revenue |
| To licensed product (M1–M7) | 18–24 weeks | ~$200–2,400 | $46–86K | General availability, self-serve, automated billing |
What runs in parallel
M2 → M6 → M7 → M9
Sequential build. Each milestone depends on the previous. This is the critical path. Cannot be parallelised internally.
M4 → M8 (parallel)
Completely independent of infrastructure. Start recruiting Week 2. Profiling sessions run while engineers build auth and AWS. Profile activates Week 8–12 regardless of infrastructure state.
M1 → M5 (parallel)
Corpus rehab (M1) is immediate. Automated research and preference feedback (M5) can follow independently. Both improve output quality for any customer at any point.
Revenue ramp by milestone
| Milestone reached | Timeline | Customers | Monthly revenue range | Gross margin |
|---|---|---|---|---|
| M3 — First pilot | Week 6 | 1 | $500–2,000/mo | ~90%+ |
| M4 — Premium add-on | Week 12 | 1 premium | $1,500–5,000/mo | ~85% |
| M6 — Customers 2+3 | Week 16 | 3 | $3,000–9,000/mo | ~90% |
| M7 — General licensing | Month 6 | 5–15 | $5,000–30,000/mo | ~94% |
| M8 — Premium network live | Month 6–8 | 5 base + 3 premium | $10,000–50,000/mo | ~93% |
| M10 — Empirical layer | Month 9–12 | 10–30 | $15,000–75,000/mo | ~95% |
| M12 — Enterprise | Month 12+ | 1 enterprise + base | $30,000–100,000/mo | ~95% |
Do first — blocks everything
M1: Corpus rehab (Week 1)
ImmediateZero cost. Massive quality improvement. Do this before anything else — every subsequent case benefits.
M4: Start strategist recruitment (Week 2)
Cannot deferProfiling takes 4–6 weeks minimum and cannot be compressed. Every week you delay recruiting is a week later the premium tier activates. Start now, run in parallel.
M2: Containerise (Week 2–4)
Infrastructure foundationNothing can be reliably cloud-deployed without Docker. All subsequent infrastructure milestones depend on this.
M3: First customer (Week 3–6)
Revenue + feedbackReal customer feedback before M6 and M7 will materially change what you build. Do not over-engineer before you have validated what customers actually want.
Key risks
Strategist quality and willingness
High riskM4 and M8 are relationship challenges, not engineering ones. The premium tier is only as strong as the strategists who agree to participate and complete profiling. Start early, expect slow recruitment.
Skipping M3 and building M7 first
Critical failure modeThe most likely mistake: building the full licensed platform before validating that anyone will pay. Get one customer at Week 6. Use that revenue and feedback to inform M7.
ABS/polling data complexity (M10)
Medium riskReal data sources are always messier than expected. M10 is Phase 4 — do not let it block M7. It can be launched independently as an intelligence upgrade after the platform is live.
DGX single point of failure
Medium riskUntil M9 (hardening), a DGX failure takes all inference offline. Acceptable for pilot customers; not for enterprise. Colocation or cloud GPU fallback added during M9.