Skip to content

Cost Estimates — Phase 1

Topline Spend Buckets

Below are the major spend buckets we've estimated for Phase 1. This is intentionally “at-a-glance” so an investor can see where burn is going before diving into assumptions.

Monthly run-rate (Early Phase 1)

  • Founders (base): $33,333/mo
  • Founders (burden line item, 10%): $3,333/mo
  • Core infra: $1,000/mo
  • Support vendors: $500/mo
  • Trust & safety vendors: $200/mo
  • Paid ads (Reddit + LinkedIn): $2,500/mo
  • Reddit (candidates): $2,000/mo
  • LinkedIn (employers, Lead Gen + retargeting): $500/mo
  • Variable compute (LLM inference + embeddings): ~$742/mo (LLM dominates; embeddings negligible)

Early Phase 1 total monthly burn (run-rate): ~$41,608/mo

Monthly run-rate (Late Phase 1 — ops contractor for 6 months)

  • Everything above, plus:
  • Ops support contractor (annual-equivalent $70k / 12): $5,833/mo (for 6 months)
  • Ops support contractor burden cushion (10%): $583/mo (for 6 months)

Late Phase 1 total monthly burn (run-rate): ~$48,024/mo

One-time Phase 1 costs: 100,000

  • Design contractors: $50,000
  • Legal (conservative): $30,000
  • Grab bag (one-time + unknown unknowns): $20,000

TL;DR

  • Phase 1 Green is a people-burn story. Fixed monthly burn dwarfs token and ad spends
  • LLM inference is ~\$742/mo and embeddings are ~\$0.22/mo (effectively zero).
  • Adding always-on paid:
    • Reddit: $2,000/mo aimed at candidate inventory
    • LinkedIn: $500/mo aimed at employer lead gen (Lead Gen Forms + retargeting)
  • Estimated monthly burn:
    • Early Phase 1: ~$41.6k/mo
    • Late Phase 1 (6 months with ops contractor): ~$48.0k/mo
  • One-time costs modeled: $100k (design + legal + unknown-unknowns).

Scope & Ground Rules

Snapshot

  • Snapshot: Phase 1 Green

What this estimate includes

  • Founders, contractors, core infra, vendors, trust & safety vendors, legal, contingency (“unknown unknowns”)
  • Paid acquisition budgets (Reddit + LinkedIn)
  • Token-driven LLM inference cost (mainline + follow-up loop)
  • Embedding generation API cost (facet summaries → embeddings)

What this estimate explicitly excludes

  • Qdrant hosting/storage (assumed negligible vs tokens at Phase 1)
  • WorkOS SSO (pass-through per connection, rolled into B2B deals)
  • Benefits/healthcare (explicitly not included; founders paid directly; ops hire contractor-based)

Phase 1 Green Usage Model (the driver for variable compute)

Agent counts

  • Candidate agents: 3,000
  • Company/role agents: 600
  • Total agents: 3,600

Traffic & engagement assumptions (Phase 1 Green baseline)

  • Weekly non-owner visitors per agent: 3
  • Engagement rate (visitors who submit ≥1 prompt): 25%
  • Prompts per engaged visitor: 3

Derived (weekly): - Non-owner visits/week: 3,600 × 3 = 10,800 - Engaged visitors/week: 10,800 × 0.25 = 2,700 - Prompts/week: 2,700 × 3 = 8,100 - Avg prompts per visit: 8,100 / 10,800 = 0.75 - Prompts per agent/week: 8,100 / 3,600 = 2.25

LLM Inference: Call Structure & Follow-up Loop

Mainline call structure

  • Assume: 3 LLM calls per prompt on average
  • Example decomposition: synthesis + PII/safety + claim/update (or similar)

Follow-up loop (separate + additive)

  • Follow-up trigger rate: 25% of prompts
  • Follow-up depth: 1.5 questions per trigger (questions ≠ calls)
  • Implementation: 1 LLM call per trigger (generates the set of follow-up questions)

Derived (weekly): - Triggers/week: 8,100 × 0.25 = 2,025 - Follow-up questions/week: 2,025 × 1.5 = 3,038 (questions) - Follow-up calls/week: 2,025 (calls)

Token Budget Assumptions

Per call budget (starting baseline)

  • Input tokens/call: 1,000
  • Output tokens/call: 400
  • Total tokens/call: 1,400

Weekly call totals

  • Mainline calls/week: 8,100 prompts × 3 calls/prompt = 24,300
  • Follow-up calls/week: 2,025
  • Total calls/week: 24,300 + 2,025 = 26,325

Weekly token totals

  • Mainline tokens/week: 24,300 × 1,400 = 34,020,000
  • Follow-up tokens/week: 2,025 × 1,400 = 2,835,000
  • All-in tokens/week: 36,855,000

Breakdown: - Input tokens/week: 26,325 calls × 1,000 = 26,325,000 - Output tokens/week: 26,325 calls × 400 = 10,530,000

Monthly token totals

Using 52/12 ≈ 4.333 weeks/month: - Input tokens/month: 26.325M × 4.333 ≈ 114.08M - Output tokens/month: 10.53M × 4.333 ≈ 45.63M

Tokens per asker prompt (useful for unit economics)

  • Follow-up calls per prompt = 2,025 / 8,100 = 0.25
  • Avg calls per prompt (including follow-up loop) = 3 + 0.25 = 3.25

So: - Input tokens per prompt: 3.25 × 1,000 = 3,250 - Output tokens per prompt: 3.25 × 400 = 1,300

Variable Cost: LLM Inference (GPT-4o)

Pricing used (assumption for estimating)

  • GPT-4o: $2.50 / 1M input tokens, $10.00 / 1M output tokens

Monthly inference cost (at Green)

  • Input: 114.08 × $2.50 = $285.20 / month
  • Output: 45.63 × $10.00 = $456.30 / month
  • Total inference cost: $741.50 / month (≈ $742/month)

Variable Cost: Embedding Generation (Facet Summaries → Embeddings)

Your embedding approach (summary)

  • Embed summaries derived from capability claims (not raw text)
  • Maintain separate semantic facets:
  • skills, experience, landscape, intent, mobility
  • Claims grouped by facet → summarized together → summary embedded and stored in Qdrant
  • Qdrant storage spend excluded from this model

Embedding parameters you set

  • Embedding dimensions: 3,072
  • Facets: 5
  • Avg facet summary length: 200 tokens
  • Facet refreshes per resolved trigger: 1.2
  • Resolution rate: 80% of triggers resolve
  • Embedding cost assumption: $0.13 / 1M tokens (for “large” embeddings)

One-time initial embedding cost

Vectors to seed: - 3,600 agents × 5 facets = 18,000 facet vectors

Tokens to embed once: - Tokens per agent (all facets) = 5 × 200 = 1,000 - Total initial tokens = 3,600 × 1,000 = 3.6M tokens

Cost: - 3.6 × $0.13 = $0.47 one-time

Ongoing monthly re-embedding cost

Weekly resolved triggers: - Resolved triggers/week = 2,025 × 0.80 = 1,620

Facet updates/week: - Facet refreshes/week = 1,620 × 1.2 = 1,944

Embedding tokens/week: - 1,944 × 200 = 388,800 tokens/week

Embedding tokens/month: - 388,800 × 4.333 ≈ 1.685M tokens/month

Cost/month: - 1.685 × $0.13 = $0.219/month (≈ $0.22/month)

Conclusion: Embeddings are functionally negligible at Phase 1 Green scale.

Guiding strategy (channel split)

  • Reddit: primary channel for candidate inventory generation (always-on, prospecting-heavy; retargeting if/when useful)
  • LinkedIn: primary channel for employer pipeline capture (retargeting + Lead Gen Forms; not cold prospecting)

Reddit (candidates)

  • Objective: candidate profile starts (agent_created)
  • Budget: $2,000/month (always-on)
  • Approach: subreddit-targeted prospecting; add retargeting later if it improves quality/CPA

LinkedIn (employers)

  • Primary objective: employer lead capture (Lead Gen Form + retargeting), not cold clicks
  • Budget: $500/month (always-on)
  • Lead definition: Work email + company + role + “hiring in next 12 months” (checkbox)
  • Target: 5 employer leads/month (implied working CPL target: ~$100/lead)

Notes on goals vs budgets

  • The LinkedIn budget is intentionally sized to the small goal (5/mo) and to avoid “LinkedIn becomes the whole burn.”
  • Reddit remains the main spend because the candidate funnel is volume-driven and you want iteration without LinkedIn CPC economics.

Fixed Costs (People + Operating)

Payroll / compensation posture

  • Founders paid directly (no benefits)
  • Ops support is contractor-based (modeled as annual-equivalent + cushion)
  • No benefits
  • Accounting for employment “burden” will occur in Phase 2, when people spend necessitates full payroll

Cost inputs

  • Founders: 2 × $200k base salary/year
  • Design contractors: $50k lump sum (Phase 1)
  • Core infra: $1,000/month
  • Support vendors: $500/month
  • Trust & safety vendors: $200/month (founder-heavy; inexpensive vendors)
  • Legal: $30k lump sum (explicitly conservative)
  • Grab bag: $20k lump sum (“one-time + unknown unknowns”)
  • Ops support: contractor equivalent of $70k annual for 6 months
  • WorkOS SSO: pass-through per connection (excluded from Phase 1 model)

Fixed cost table (fixed vs timing)

Cost area Fixed/Variable Timing Assumption Monthly run-rate One-time
Founders (base) Fixed Phase 1 2 × $200k base $33,333 —
Founders (burden) Fixed Phase 1 10% overhead (no benefits) $3,333 —
Design contractors Fixed Phase 1 Lump sum — $50,000
Core infra Fixed Phase 1 Self-managed $1,000 —
Support vendors Fixed Phase 1 Workspace/QBO/subscriptions $500 —
Trust & safety vendors Fixed Phase 1 Cheap vendor stack $200 —
Paid ads (Reddit + LinkedIn) Fixed Phase 1 Always-on budgets $2,500 —
Legal (conservative) Fixed Phase 1 Mostly boilerplate + cleanup — $30,000
Grab bag Fixed Phase 1 One-time + unknown unknowns — $20,000
Ops support (contractor) base Fixed Late Phase 1 (6 mo) $70k annual-equivalent / 12 $5,833* —
Ops support cushion Fixed Late Phase 1 (6 mo) +10% cushion $583* —

* Ops contractor line applies only during the 6 months you include the ops role.

Fixed monthly burn

  • Early Phase 1 (no ops contractor yet):
  • $33,333 + $3,333 + $1,000 + $500 + $200 + $2,500 = $40,866/month
  • Late Phase 1 (with ops contractor for 6 months):
  • $40,866 + $5,833 + $583 = $47,282/month

Fixed one-time costs (Phase 1)

  • Design $50,000 + Legal $30,000 + Grab bag $20,000 = $100,000

Combined Burn: Fixed + Variable (Phase 1 Green)

Variable monthly (usage-driven)

  • LLM inference: ~$742/month
  • Embedding generation: ~$0.22/month (negligible)

Total variable compute: ~$742/month

Total monthly burn (run-rate)

  • Early Phase 1: $40,866 fixed + $742 variable ≈ $41,608/month
  • Late Phase 1 (6 months w/ ops contractor): $47,282 fixed + $742 variable ≈ $48,024/month

One-time Phase 1 costs

  • $100,000 (design + legal + unknown-unknowns)

Unit Economics (Investors Will Ask)

Compute cost per asker prompt

Per prompt: - Input tokens ≈ 3,250 → 0.00325M × $2.50 = $0.00813 - Output tokens ≈ 1,300 → 0.00130M × $10.00 = $0.01300 - Total ≈ $0.0211 per prompt (~2.1¢)

Compute cost per engaged session

  • Prompts per engaged visitor = 3
  • Cost ≈ 3 × $0.0211 = $0.0633 (~6.3¢ per engaged session)

Compute cost per non-owner visit

  • Avg prompts per visit = 0.75
  • Cost ≈ 0.75 × $0.0211 = $0.0158 (~1.6¢ per visit)

Compute cost per agent-week

  • Prompts per agent-week = 2.25
  • Cost ≈ 2.25 × $0.0211 = $0.0475 (~4.8¢ per agent-week)

These are planning targets for reporting and iteration; they are not “guarantees.”

  • LinkedIn employer CPL target: ~$100/lead
  • Lead = work email + company + role + hiring-in-12-months checkbox

  • Goal = 5 leads/mo on $500/mo budget

  • Reddit candidate CPA: use actuals and iterate

  • Optimize to agent_created but report quality-adjusted conversion (e.g., agent_activated) as you instrument it

Notes / Interpretation

Why compute is cheap at Green (in this model)

  • Prompt volume is meaningful, but token budgets and call counts are contained.
  • Even with the follow-up loop modeled as additive, inference remains under $1k/month.

Sensitivity knobs (what can blow up)

If any of these move materially, compute changes quickly: - Token budget per call (context bloat, long transcripts, verbose responses) - Calls per prompt (extra safety passes, reranking, multi-step synthesis) - Prompt depth per engaged visitor - Follow-up trigger rate / additional loops

  • Kept as a conservative lump sum ($30k). Expectation is cheaper due to boilerplate; this is intentionally padded.

Trust & safety

  • Modeled as $200/month for inexpensive vendors + founder effort (explicitly not staffing-heavy yet).

Ops support

  • Pushed to late Phase 1, contractor-based, modeled for 6 months only.

Quick Checklist (to be updated later with real data)

These will dial in as real telemetry comes in: - Actual prompts per engaged visitor - Actual token usage per call (measured from logs) - Actual calls per prompt (instrument each model invocation) - Actual follow-up trigger rate and resolution rate - Actual facet summary length (tokens) and facet refreshes per resolution - Actual CPL on LinkedIn for employer leads (and lead quality) - Actual Reddit CPA for candidate starts, plus activation/completion rates