For Investors

AI & NLT Platform

PitchPad is the proof case for a portfolio rule: every venture needs product-AI (the customer feels it) and delivery-AI (NLT’s platform shrinks research, ship, and ops cost). One without the other fails the thesis — a clever model with a four-person eng crew, or a cheap ship loop with no AI in the product, both lose to incumbents who can outspend on either axis.

Both required

Product-AI: On-porch voice → structured scope → live rate-card pricing. The AI is the close motion, not a chatbot bolted onto a dispatcher desk. Without it, PitchPad is just another field-service CRM.
Delivery-AI: NLT Memory, Research Library, AgentForge workflows, Claude/Cursor seats, and LiteLLM are how a 1–2 person product team ships research briefs, investor surfaces, smokes, and MVP scaffolding while Jobber- class competitors staff those loops with people. Modeled Modeled — headcount vs incumbent staffed research/ship/QA loops; same Phase 1–3 staffing tables as Path to Product.

Product-AI — what the operator gets

Category leaders built around the dispatcher desk. PitchPad inverts that: the truck is the primary surface. AI sits in the walk-through itself.

On-device voice inference — Llama 3.2 1B (quantized) on WebGPU with Whisper-server fallback when confidence drops below 0.78. Scope capture stays on the porch even with weak cellular. Modeled Modeled — internal curb-latency benchmark; 240 sample utterances across four categories on iPhone 15 / Pixel 8 (also cited on Market Research).
Category-tuned rate card — voice + photo + modifiers map into priced line items for cleaning, painting, pest, and lawn. The model is the pricing edge; the QuickBooks / Stripe handoff is the settle.
Porch close — proposal, e-signature, and deposit in one continuous motion so warm-lead decay never leaves the curb. Sourced Sourced — warm-lead decay framing from Thumbtack / HomeAdvisor field-services response studies (see Market Research find). Source ↗

Architecture, integrations, and tech risks live on Engineering. This tab owns the investor thesis that the product is AI-core, not AI-wrapper.

Delivery-AI — the NLT stack (not just AgentForge)

“NLT Platform” is a mix. AgentForge is the runtime; Memory and the Research Library are the compounding knowledge layer; seats and LiteLLM are the human+agent build surface; quality tooling protects ship gates.

Layer	What PitchPad uses it for	Who owns it
NLT Memory	FUND/PASS lessons, role feedback, brand patterns recalled before research and ship agents write.	Shared (tapps-brain / bridge)
Research Library	Pre-scored BLS / FTC / contractor sources with citation provenance — preferred over cold web search.	Shared (NLT Labs)
AgentForge	`pe-evaluate`, `poc-ship`, `mvp-ship`, smoke harnesses, dossier regen — publish in, do not rebuild.	Shared runtime
Claude / Cursor seats	Product engineers implement PWA + integrations with agent-assisted loops; budgeted opex, not silent founder cards.	Product eng + NLT
LiteLLM gateway	Routed Gemini/Claude inference for agent ship runs and product voice fallback when on-device misses.	Shared gateway
TAPPS quality gates	Score / gate / security / content-QA on Python and POC surfaces so smokes stay reinforced without a QA hire.	Shared tooling

Sourced Sourced — stack inventory from live NLT workflows in-repo (pe-evaluate, poc-ship, mvp-ship) and Memory / Research Library runbooks; not a vendor brochure claim.

Lifecycle — how AI implements PitchPad

Each row names the delivery-AI mechanism and the human gate. Product-AI appears again at MVP when the operator-facing loop ships.

Stage	Delivery-AI	Product-AI touch	Human gate
Research	Memory recall + Research Library + Exa/Firecrawl agents in `pe-evaluate`	Scope checks that voice/pricing AI is core, not optional	Creative Director brief
Evaluate	Investment Committee + red team + buyer sim on AgentForge	Execution score assumes capital-efficient AI build	Bill FUND / PASS
POC	`poc-ship` agents: design, copy, content-QA, builder → this /inside brief	Demo narrative makes the porch-close AI concrete	Content / brand lints
MVP	`mvp-ship` + reinforced smokes; scaffolds publish into AF	Live voice → proposal → deposit on app.pitchpad	Operator concierge + Stripe underwriting
Operate	Memory writes from FUND/PASS and ship outcomes; LiteLLM spend in eng envelopes	Per-category prompt tuning during concierge	Founding engineer + Bill

Sourced Sourced — stage names and agent roles match workflows/pe-evaluate.yaml, poc-ship.yaml, mvp-ship.yaml and the Process tab DAG for this ship.

This-run agent timings and costs: see Process.

Cost to compete

Incumbents win with sales teams and multi-year product organizations. PitchPad’s wager is that product-AI (porch close) plus delivery-AI (shared NLT stack) keep eng at 1–2 FTE through Phase 2 and 2 FTE in Phase 3 instead of 4–5 product engineers plus a platform hire. Phase 3 eng modeled at $240K on the platform path versus roughly $480–620K people+tools on a DIY agent stack over the same window.

Modeled Modeled — Path to Product Phase 3: $240K = 2 product engineers × ~10 months at founder/seed cash rates; DIY counterfactual ≈ 4 eng + platform/DevOps hire over the same window ($480–620K people+tools). Not a GAAP forecast.

Phase cash uses and rollback triggers: Path to Product. The numbers there treat Claude/Cursor seats and LiteLLM as explicit opex — delivery-AI is paid for, not assumed free.

Memory flywheel

Delivery-AI compounds. Each pe-evaluate and poc-ship run recalls prior FUND/PASS patterns and role feedback before writing; outcomes write back to hive/project memory with durable tags. PitchPad inherits lessons from sibling ships (citation provenance, brand lint, Concierge MVP kinematics) instead of relearning them with another PM hire. That is why the platform mix matters: AgentForge alone orchestrates; Memory is what makes the next evaluation cheaper and less wrong.

Sourced Sourced — recall-before-write and hive FUND/PASS tags are mandatory in PE agent prompts and docs/workflows/pe/MEMORY-INSTRUCTIONS.md.

Platform risks (honest)

Risk	Mitigation	Evidence
Shared AgentForge outage blocks evaluate/ship automation	Product code and Render deploy remain manual-operable; AF is speed, not sole production host for the PWA	Modeled Modeled — separation of AF orchestration vs apps/pitchpad + Render hosting in the ship architecture (MVP path).
Inference / seat spend scales faster than plan	Line-itemed in Path to Product; execution_caps on pe-evaluate economy profiles for research	Sourced Sourced — seat + LiteLLM line items on Path to Product; pe-evaluate execution_caps / economy profile in docs/workflows/pe/EXECUTION_PROFILE.md.
On-device product-AI accuracy under bad audio	Whisper fallback + concierge prompt tuning — see Engineering T1	Modeled Modeled — Engineering tech-risk T1; 0.78 confidence threshold + server Whisper fallback path.
Memory pollution from bad FUND/PASS feedback	Tiered scopes, citation provenance lint, founder retro tags before hive write	Assumption Assumption — controls exist in runbooks; no measured pollution rate on PitchPad yet.