// P03 — B2B Sales Intelligence Agent // Planner-Executor-Reflector loop · Tavily + HN · personalized outreach // ── Terminal demo: 72s ────────────────────────────────────────────────── const P03_TERMINAL_SCRIPT = [ { t: 0.3, kind: 'caption', text: '01 / Clone the repository' }, { t: 0.8, kind: 'cmd', text: 'git clone git@github.com:Juadsuarezsan/sales-intelligence-agent.git', cwd: '~' }, { t: 2.6, kind: 'out', text: "Cloning into 'sales-intelligence-agent'..." }, { t: 2.95, kind: 'out', text: "remote: Enumerating objects: 156, done." }, { t: 3.25, kind: 'out', text: "Receiving objects: 100% (156/156), 268.7 KiB | 7.1 MiB/s, done." }, { t: 3.55, kind: 'out', text: "Resolving deltas: 100% (54/54), done." }, { t: 4.05, kind: 'cmd', text: 'cd sales-intelligence-agent', cwd: '~' }, { t: 5.4, kind: 'caption', text: '02 / Configure API keys' }, { t: 5.8, kind: 'cmd', text: 'cp .env.example .env' }, { t: 6.8, kind: 'cmd', text: '$EDITOR .env' }, { t: 7.6, kind: 'editor', title: '.env', lines: [ "ANTHROPIC_API_KEY=sk-ant-•••••••••••••••••", "ANTHROPIC_MODEL=claude-sonnet-4-5", "", "# Web search (1000 free / month, $0.001 after)", "TAVILY_API_KEY=tvly-•••••••••", "", "# HN Algolia is free, no key needed", "DATABASE_URL=postgresql://sales:dev@localhost:5432/sales", "", "# Agent loop budget", "MAX_TOOL_CALLS_PER_COMPANY=8", "MAX_REFLECTION_ITERATIONS=3" ] }, { t: 11.6, kind: 'out', text: '✓ .env saved · 0 secrets staged · API keys configured', color: 'success' }, { t: 13.0, kind: 'caption', text: '03 / Bring up stack + install deps' }, { t: 13.4, kind: 'cmd', text: 'docker compose up -d postgres' }, { t: 14.0, kind: 'out', text: ' ✔ postgres-1 pgvector/pgvector:pg16 healthy', color: 'success' }, { t: 14.4, kind: 'cmd', text: 'uv pip install -e ".[dev]"' }, { t: 15.0, kind: 'progress', start: 15.0, end: 17.6, total: 124, unit: 'packages', label: 'Installing' }, { t: 17.8, kind: 'out', text: 'Installed 124 packages in 2.41s', color: 'success' }, { t: 18.0, kind: 'out', text: ' + tavily-python==0.5.0 selectolax==0.3.27 httpx==0.27.2', color: 'accent' }, { t: 18.2, kind: 'out', text: ' + langgraph==0.2.45 pydantic==2.9.2 anthropic==0.39.0', color: 'accent' }, { t: 18.8, kind: 'caption', text: '04 / Load YC companies dataset (5,124 rows)' }, { t: 19.2, kind: 'cmd', text: 'python -m src.data.load_yc' }, { t: 20.0, kind: 'out', text: '→ Pulling yc-oss/yc-companies @ HEAD' }, { t: 20.6, kind: 'progress', start: 20.6, end: 22.6, total: 5124, unit: 'companies', label: 'Loading' }, { t: 22.8, kind: 'out', text: '✓ Loaded 5,124 YC companies (id, name, batch, industry, status, founded, team_size, website)', color: 'success' }, { t: 23.1, kind: 'out', text: '→ Sampled 100 active companies for batch processing (seed=42)', color: 'accent' }, { t: 23.8, kind: 'caption', text: '05 / Run plan-execute-reflect loop on 100 companies (parallel, 8 workers)' }, { t: 24.2, kind: 'cmd', text: 'python -m src.agents.run_batch --n 100 --workers 8' }, { t: 25.0, kind: 'out', text: 'INFO [batch] Spawning 8 worker tasks · max 3 reflection loops per company' }, { t: 25.4, kind: 'out', text: 'INFO [batch] Each worker: Planner → Executor → Reflector → ProfileBuilder → EmailWriter' }, { t: 25.7, kind: 'progress', start: 25.7, end: 36.4, total: 100, unit: 'companies', label: 'Processing' }, { t: 36.6, kind: 'out', text: '✓ 100 companies processed in 10.7 min · 12.3 avg tool calls/company', color: 'success' }, { t: 36.9, kind: 'out', text: ' ├─ Tavily calls: 1,218 · cache hit 14%', color: 'accent' }, { t: 37.1, kind: 'out', text: ' ├─ HN search calls: 428', color: 'accent' }, { t: 37.3, kind: 'out', text: ' ├─ Website fetches: 294', color: 'accent' }, { t: 37.5, kind: 'out', text: ' └─ Reflection loops: 213 · avg 2.13 per company', color: 'accent' }, { t: 38.4, kind: 'caption', text: '06 / Quality scoring with Claude Opus-as-judge' }, { t: 38.8, kind: 'cmd', text: 'python -m src.eval.score_outputs --judge claude-opus-4' }, { t: 39.6, kind: 'out', text: 'INFO [eval] Rubric: personalization (1-5) · accuracy (1-5) · CTA clarity (1-5)' }, { t: 40.0, kind: 'progress', start: 40.0, end: 47.0, total: 300, unit: 'evals', label: 'Judging' }, { t: 47.2, kind: 'table', cols: ['system', 'accuracy', 'personalization', 'CTA clarity', 'p95', '$/company'], rows: [ ['Template only', '—', '1.4', '4.8', '2.0s', '$0.001'], ['Single-search + email', '3.1', '2.7', '3.9', '8.2s', '$0.012'], ['Agent loop (this system)', '4.4', '4.6', '4.5', '18.4s', '$0.041', 'highlight'], ]}, { t: 55.0, kind: 'out', text: '→ +1.9 personalization score vs single-search baseline', color: 'success' }, { t: 56.0, kind: 'caption', text: '07 / Persist 100 enriched profiles + drafts to Postgres' }, { t: 56.4, kind: 'cmd', text: 'python -m src.persist --commit' }, { t: 57.4, kind: 'out', text: 'INSERT INTO company_profile ... 100 rows OK' }, { t: 57.8, kind: 'out', text: 'INSERT INTO outreach_email ... 100 rows OK' }, { t: 58.2, kind: 'out', text: 'INSERT INTO research_log ... 1,940 rows OK' }, { t: 58.5, kind: 'out', text: '✓ DB tx committed · 100 companies queryable via /api/profiles', color: 'success' }, { t: 58.9, kind: 'out', text: '→ Traces pushed to LangSmith (30 agent loops public · others private)', color: 'accent' }, { t: 59.8, kind: 'caption', text: '08 / Launch Next.js gallery (Tailwind + shadcn)' }, { t: 60.2, kind: 'cmd', text: 'cd frontend && pnpm dev' }, { t: 61.0, kind: 'out', text: ' ▲ Next.js 14.2.15 → ready in 1.3s' }, { t: 61.4, kind: 'out', text: ' ✓ Local: http://localhost:3000', color: 'accent' }, { t: 61.8, kind: 'out', text: ' ✓ Gallery: 100 companies pre-loaded, click any to see the research timeline', color: 'success' }, { t: 62.6, kind: 'caption', text: '✓ Agent gallery is live — see Demo 02 →' }, ]; const P03TerminalDemo = makeTerminalDemo({ script: P03_TERMINAL_SCRIPT, duration: 64, persistKey: 'p03-terminal', title: 'zsh — sales-intelligence-agent — 132×42', cwd: '~/sales-intelligence-agent', captionPrefix: 'P03', }); // ── System demo: agent loop visualization ──────────────────────────────── // Two companies processed end-to-end (~75s) // Layout: // Top bar (60px) // Left pane (research log + tool calls, 380px) // Center pane (CompanyProfile being built, 360px) // Right pane (outreach email being composed, rest) const P03_TIMELINE = { companies: [ { start: 1, name: 'Anthropic', meta: 'YC W21 · AI · 1k+ team · San Francisco · Founded 2021', website: 'anthropic.com', // Tool calls along the way tools: [ { t: 5.0, type: 'tavily', q: '"Anthropic" recent funding 2025', hits: 8 }, { t: 7.4, type: 'website', q: 'anthropic.com/news (sitemap)', hits: 14 }, { t: 9.2, type: 'hn', q: 'Anthropic Claude 4.5', hits: 22 }, { t: 11.6, type: 'reflect', q: 'enough info? gaps: revenue model, hiring signals', dec: 'refine' }, { t: 13.4, type: 'tavily', q: '"Anthropic" job postings December 2025', hits: 11 }, { t: 15.6, type: 'reflect', q: 'gaps closed · proceed to profile', dec: 'proceed' }, ], // Profile fields with reveal times profile: [ { t: 9.0, k: 'name', v: 'Anthropic, PBC' }, { t: 10.5, k: 'industry', v: 'AI safety + LLM platform' }, { t: 12.0, k: 'team_size', v: '1,200+ (Dec 2025)' }, { t: 14.0, k: 'recent_funding', v: 'Series D · $8B led by Amazon (Nov 2024)' }, { t: 16.0, k: 'hiring_signal', v: 'AI Safety + ML Engineer · ~340 open' }, { t: 17.5, k: 'pain_point', v: 'Scaling enterprise eval infra for safety tier' }, ], // Email composition (streamed chars) email: { startStream: 18.5, endStream: 23.5, subject: 'Cut your safety-eval pipeline time by 70% (specific to Claude 4.5)', body: 'Hi {first_name},\n\n' + 'Saw the Claude 4.5 launch + your 340 open ML/safety roles — your eval throughput\n' + 'is about to become the bottleneck the moment hiring closes.\n\n' + 'I just shipped a portfolio project that runs RAGAS over 100 queries in 7 minutes\n' + 'against 3 vector DBs in parallel. The pattern transfers cleanly to your safety-tier\n' + 'eval set — happy to share the architecture diagram, no pitch.\n\n' + 'Worth 15 minutes next week?\n— Juan', scores: { personalization: 4.6, accuracy: 4.8, cta: 4.5 }, }, }, { start: 35, name: 'Linear', meta: 'YC S19 · SaaS · 120 team · San Francisco · Founded 2019', website: 'linear.app', tools: [ { t: 37.0, type: 'tavily', q: '"Linear" Series C funding', hits: 6 }, { t: 39.0, type: 'website', q: 'linear.app/changelog', hits: 9 }, { t: 41.0, type: 'hn', q: 'Linear AI Cycle planning', hits: 14 }, { t: 43.0, type: 'reflect', q: 'sufficient · proceed to profile', dec: 'proceed' }, ], profile: [ { t: 41.5, k: 'name', v: 'Linear, Inc' }, { t: 42.5, k: 'industry', v: 'Project management for engineering' }, { t: 43.5, k: 'team_size', v: '120 (Q4 2025)' }, { t: 44.5, k: 'recent_funding', v: 'Series C · $200M @ $1.25B (Jun 2024)' }, { t: 45.5, k: 'pain_point', v: 'Onboarding for non-eng teams · Asana defection retention' }, ], email: { startStream: 46.5, endStream: 51.5, subject: 'Reduce non-eng onboarding from 2 weeks to 2 days', body: 'Hi {first_name},\n\n' + 'Picked Linear over Asana 6 months ago. The wall I hit: every non-engineering\n' + 'colleague needs hand-holding on cycle ceremony for ~2 weeks before they\'re fluent.\n\n' + 'I prototyped an onboarding agent for a similar workflow tool. It reduces day-1\n' + 'time-to-first-issue from 35 min to 4 min, measured. Built on Claude + a tiny\n' + 'retrieval layer over your docs.\n\n' + 'Open to a 20-minute walkthrough?\n— Juan', scores: { personalization: 4.5, accuracy: 4.7, cta: 4.6 }, }, }, ], }; function P03SystemDemo() { return (
); } function P03ActiveCompany() { const t = useTime(); let active = P03_TIMELINE.companies[0]; for (const c of P03_TIMELINE.companies) { if (t >= c.start) active = c; } return active; } function P03TopBar() { const t = useTime(); const active = P03ActiveCompany(); // Count loops and stats const toolCount = active.tools.filter(x => t >= x.t).length; return (
batch · agent loop · target {P03_TIMELINE.companies.findIndex(c => c === active) + 1} / 100 · {active.name}
); } function KpiPill03({ label, value, color }) { return (
{value} {label}
); } function P03ResearchLog() { const t = useTime(); const active = P03ActiveCompany(); const visibleTools = active.tools.filter(x => t >= x.t); const scrollRef = React.useRef(null); React.useEffect(() => { if (scrollRef.current) scrollRef.current.scrollTo({ top: scrollRef.current.scrollHeight, behavior: 'smooth' }); }, [visibleTools.length]); return (
research log plan → exec → reflect
{visibleTools.map((tc, i) => )} {visibleTools.length === 0 && (
▸ planner initializing...
)}
); } function P03ToolCallRow({ tool }) { const color = tool.type === 'tavily' ? '#22d3ee' : tool.type === 'website' ? '#7c5cff' : tool.type === 'hn' ? '#fbbf24' : tool.type === 'reflect' ? (tool.dec === 'proceed' ? '#34d399' : '#f472b6') : '#9aa3b8'; const label = tool.type === 'tavily' ? 'TAVILY' : tool.type === 'website' ? 'FETCH' : tool.type === 'hn' ? 'HN' : 'REFLECT'; return (
{label} {tool.hits != null && {tool.hits} hits} {tool.dec && → {tool.dec}}
{tool.q}
); } function P03ProfileBuilder() { const t = useTime(); const active = P03ActiveCompany(); const visibleFields = active.profile.filter(f => t >= f.t); return (
CompanyProfile (Pydantic) {visibleFields.length}/{active.profile.length}
{active.name}
{active.meta}
{active.profile.map((f, i) => { const visible = t >= f.t; return (
{f.k}
{visible ? f.v : '...'}
); })}
); } function P03EmailDraft() { const t = useTime(); const active = P03ActiveCompany(); const em = active.email; // Stream the email body char-by-char const allText = em.subject + '\n\n' + em.body; const elapsed = Math.max(0, t - em.startStream); const dur = em.endStream - em.startStream; const ratio = Math.min(1, elapsed / dur); const chars = Math.floor(allText.length * ratio); const streamed = allText.slice(0, chars); const showScores = t >= em.endStream; const showCursor = t >= em.startStream && t <= em.endStream; return (
outreach draft Claude · streaming
{t < em.startStream && (
▸ awaiting profile completion...
)} {t >= em.startStream && (() => { // Split subject from body const lines = streamed.split('\n'); const subjectLine = lines[0] || ''; const bodyLines = lines.slice(2); return ( <>
SUBJECT
{subjectLine}
BODY
{bodyLines.join('\n')} {showCursor && }
); })()}
{showScores && (
)}
); } function ScoreChip({ label, v }) { return (
{label}
= 4.5 ? '#34d399' : v >= 4 ? '#22d3ee' : '#fbbf24', fontSize: 18, fontWeight: 700, fontFamily: 'JetBrains Mono' }}>{v} / 5
); } Object.assign(window, { P03TerminalDemo, P03SystemDemo });