Alive pitch deck

Important: These are transparent planning scenarios for a future product, not Alive usage logs. Input/provider list prices checked October 13, 2026. They intentionally exclude speculative market-adoption claims. Recheck before external publication.

• Worker CPU/memory: DigitalOcean Basic Droplets: 1 vCPU / 2 GiB = $0.01786/hour; 4 vCPU / 8 GiB = $0.07143/hour. [P1]

• Effective productive utilization: 40% assumed for short-lived workers, with the remaining paid capacity allocated to idle time, startup, scheduling, and imperfect packing. Effective CPU-worker-hour rate = listed rate / 0.40. An actual managed job platform may bill differently.

• Budget LLM: OpenAI GPT-5.6 Luna: $0.20 / million input tokens; $1.20 / million output tokens. [P2]

• More capable LLM: OpenAI GPT-5.6 Terra: $2.00 / million input tokens; $12.00 / million output tokens. [P3]

• Token batching assumption: The A2 token volumes are totals across calls. Each request stays at or below the provider’s 272k-input-token long-context threshold; a single 500k-token Luna prompt would attract higher rates and is not priced by this base case. Cache write/read pricing is not modeled. [P2]

• Per-run storage/network/orchestration allowances: $0.015 lightweight, $0.05 standard, $0.25 advanced, assumptions, not provider invoices. Includes reasonable buffer for artifacts/network incidental usage, but not all possible vendor/tool charges.

• Persistent object storage reference: DigitalOcean Spaces starts at $5/month with 250 GiB included. [P4] Persistent baseline services should be tracked separately and not charged twice in run-level costing.

• No web-search/tool fees, dedicated GPU, third-party license fees, managed database minimums, payment processing, 24/7 dedicated workers, enterprise support labor, or compliance/legal expenses are included in per-run calculations. Unusually memory-intensive tasks would require a larger worker tier.

Formula: (worker_hours × listed_worker_hourly_rate / 0.40) + Σ[(input_tokens/1M × input_rate)+(output_tokens/1M × output_rate)] + artifact_allowance.

Interpretation: A standard numerical simulation without LLM calls could cost much less; a large agent-based evaluation with hundreds of thousands of expensive model calls could cost far more. These illustrative configurations are deliberately explicit, not universal averages.

¹ Fixed account expense is an assumed allocation toward persistent application hosting, database, observability, and storage—not an invoice or measured marginal cost. Team is one workspace, not one seat; Dev is one account. Costs and dollars are rounded only at display stage.

• Proposed credits at customer-facing billing value: Community $2, Developer $10, Team $40 each month. These are not the amount Alive necessarily spends. Unused credits expire in this illustrative setup. Paid plans have prepaid overages or pay-as-you-go limits.

• For analysis, suppose usage is valued at 2.5× underlying modeled run expense (subject to minimum charges, premium models, taxes, and future actual cost calibration): Community's sample consumes about $0.85 retail; Developer about $9.68; Team about $34.47, each within the stated allowance. This markup is a price hypothesis, not a proven market-clearing rate.

• The Developer profile costs $4.62 / $39 (about 88.1% modeled pre-personnel direct contribution margin); the Team profile costs $17.79 / $149 (about 88.1%). These are not financial-statement gross margins, and don't incorporate normal enterprise customer-success/engineering overhead.

• The Community monthly cost applies to an active free account with the specified usage. Inactive accounts have different costs; heavy users must be capped, BYOK, or moved to paid/prepaid credits. If all 1,000 free pilot users behaved like this active profile, that alone would imply about $590/month variable and allocated costs; our $400 three-month grant subsidy cannot fund that volume. Therefore the pilot must be smaller, local-first, BYOK where possible, and budget-capped.

• If hosted runs are billed at a resource rate below actual utilization-adjusted and support expense, Alive loses money on consumption. Enforce run-level preflight cost estimates, concurrency caps, reservations/refunds for credits, hard spend limits, metering, and alerts.

At $39/month with an illustrative $4.62/month account-level run + allocated expense, the remaining $34.38 can help cover shared fixed costs, customer acquisition, and other unmodeled expenses. This does not mean the company breaks even on one customer. A hypothetical $200/month in incremental shared fixed costs would require 6 similarly active Developer accounts to cover that amount under the scenario, assuming no additional unmodeled costs or team subscriptions. True break-even requires full P&L, not just marginal run costs.

Collect per-run worker-time, actual billed inference input/output, retry rates, memory profile, generated artifact bytes, cloud idle capacity, customer-tier active usage, dollar cost by scenario, and model output/replay quality. Publish internal dashboards and recompute economics monthly.

All costs below use A1–A3 without rounding intermediate calculations. Implementation: apps/site/app/deck/economics.ts; tests reconcile the source costs and revenue case.

The free subsidy is $588.59/month for 1,000 active free accounts at the stated profile. The stress case doubles utilization-adjusted worker expense and inference expense; artifact allowances and fixed account allocations remain constant. It does not double run volume. Institutional revenue and delivery cost are excluded from this contribution calculation; the separate $39,400 MRR example includes 10 institutional pilots at an assumed $500/month. These are pre-personnel direct contribution estimates, not gross margins or company profitability. Payroll, support, development, acquisition, payment fees, legal/security and other overhead must be funded from contribution. Customer counts, willingness to pay, conversion, retention, and institutional costs are unvalidated; there is no dated operating forecast or LTV/CAC claim.

Milestones

• Month 1 — Replay and explain: Publish a version-locked monitoring benchmark, boundary tests and provenance. Complete at least five builder interviews.

• Month 2 — Onboard three teams: Each completes a baseline and changed variant, trace review and documented design decision.

• Month 3 — Change and return: At least one team changes its design because of a finding; at least two return in a separate session to test a later version. Publish actual spend, outcomes and barriers.

Boundaries: No talent/payroll allocation; no assumption of 1,000+ funded cloud users; no clinical PHI collection by default; no claim of real-world clinical predictions; support only allowable grant expenditures. Reallocate within approved budget only if the sponsor allows it.

The core export contains ten grant-focused slides. The full export adds eleven seafoam appendix slides, then the complete detailed assumptions, arithmetic, grant cash plan and linked sources. The optional equity-fundraising template is excluded.

Demo reproduction and interpretation: research/demos/monitoring-stress/README.md. WorldSpec: research/demos/monitoring-stress/world.json. Evidence record: research/demos/monitoring-stress/summary.json; exact resolved models: research/demos/monitoring-stress/models.lock.json. Replication-zero event digests are replayed in automated tests. No production model, default, contract or canonical fixture changes.

All quantities are modeled_outcome from an executed mock experiment, not clinical validation. Intervals represent replication variation, not uncertainty in parameter assumptions. See the README for non-wear exclusions, censored waits, idealized comparator, fixed latency, variant changes and next tests that have not been executed.

Greenhalgh et al., Beyond Adoption, JMIR 2017, Case A: four clinical services. Young adult diabetes and cancer surgery used video for about 20% of follow-ups; prenatal diabetes abandoned it and community heart failure paused it. This is a qualitative case study, not a population-wide failure rate, and is not an Alive prediction or the founder's dataset result. Original study.

The founder's 250-intervention adoption/sustainment dataset is in development. No aggregate finding from that dataset is claimed in this deck.

AMA 2026 Physician AI Survey, March 2026, pages 4–5: 72% report incorporating at least one AI use case, versus 48% in 2024. The use-case question has 1,342 responses in 2026; 1,183 in 2024. Do not substitute 81% awareness-or-use for explicit use. Total survey participants: 1,692; denominators vary by question. Survey changes limit strict comparisons across years. Page 4 reports 88% regard safety/efficacy validation as important to adoption. These are physician self-reports, not verified deployment or Alive demand.

CMS hospital-at-home study fact sheet, September 30, 2024: program selection depends on clinical and psychosocial circumstances, home/community environment and local resource capabilities. This historical study supports the workflow argument; the deck does not claim current waiver eligibility or a regulatory requirement for Alive.

The map is editorial positioning, not measured coordinates or a claim that adjacent tools cannot support repeated tests. Simulation packages can automate experiments. Synthea spans repeated generation and study inputs, so appears in both left quadrants. The right-upper cell describes Alive's intended product workflow; the status slide distinguishes implemented local experiments from planned integrations.

• Antithesis: reproducible software simulation testing.

• Synthea: open-source synthetic patient and record generation; a complementary fixture source, not an operational workflow twin.

• AnyLogic: multi-method simulation.

• Simul8: healthcare process simulation.

• FlexSim Healthcare: operational simulation and capacity studies.

• Synthetic Users: AI persona research.

• ORCHA: digital-health assurance.

Historical list-price inputs, checked October 13, 2026. Recheck before relying on them; calculation precision is not measurement accuracy.

• [P1] DigitalOcean Basic Droplets.

• [P2] OpenAI GPT-5.6 Luna.

• [P3] OpenAI GPT-5.6 Terra.

• [P4] DigitalOcean Spaces.

Recruitment and success counts are proposed targets, not signed partners or traction. Five discovery interviews precede three participating teams. Completion means a baseline and changed variant, a recorded trace review and a design decision. A changed design must link to an experiment finding. A return means a separate session testing a later version by month three, not another click during onboarding.

Budget: $800 for evaluation inference and permitted evaluation resources; $450 simulation compute and tests; $400 capped pilot usage credits; $250 hosting/database/storage; $100 contingency. Resource purchases depend on sponsor eligibility and demonstrated need. No contractor or personnel compensation is budgeted. Record receipts and metered consumption; unused credits are not automatically spend. Stop onboarding or cloud runs before exceeding the cap.

Founder weekly availability, recruitment-role compensation and any additional contributors remain unconfirmed. Neither interviewing candidates nor past institutional work is counted as pilot staffing. The pilot must have a founder-owned schedule before launch; optional hires are outside its critical path.

L1–L4 are internal terms scoped to a metric, population and use. L1: execution with traces and assertions. L2: internal replay, conservation, boundary and distribution tests. L3: calibration to documented external data with tolerances and applicability. L4: independent retrospective or prospective validation with disclosed predictive errors. Evidence promotion requires a versioned report and reviewer sign-off; no model grade is promoted by this deck.

Choose a published operational comparator with usable data and rights, declare outputs and tolerances before fitting, retain a holdout, report discrepancies and sensitive assumptions, then obtain independent review. A back-test against a known implementation outcome is useful only when the modeled mechanisms and population fit. The grant does not promise clinical validation or patient benefit.