🇵🇹 Portugal Tech Week • 6–15 NOV 2026 • Hands-On Workshop

Building Vision • Shaping The Future

Build a Multimodal Pitch & Deck Critic
with Gemma 4 & Local Ollama

Turn your laptop into a ruthless, 100% private AI Investment Committee that listens to a 30-second spoken elevator pitch (16kHz WAV), inspects slide deck visuals (280 vs 1,120 vision tokens), catches cross-modal contradictions, and verifies unit economics with deterministic tools.

Native Audio Ingestion

Feed raw 16kHz 16-bit mono PCM WAV audio directly into gemma4:e2b with zero external Whisper ASR step.

Variable Vision Budgets

Compare 280 tokens (448×448 layout) vs. 1,120 tokens (1120×1120 dense financial tables & 8pt footnotes).

Deterministic VC Tools

4 native Ollama function calls (audit_unit_economics, verify_tam_math, density & buzzword checks).

⚠️ WORKSHOP GUIDE DISCLAIMER: This is an educational workshop and is just a guide to better pitch your project — this isn't the real result.

100% Local • Zero Cloud Retention

👤 Speaker: Nuno João Andrade • Quick Pre-Flight Check:

👋 Meet Your Workshop Instructor

Nuno João Andrade

Senior Developer & Cloud Platform Expert (GCP / AWS) • Google Developer Expert (GDE) for Cloud

Nuno João Andrade

Google Developer Expert • Cloud & AI

🔗 LINKEDIN PROFILE

www.linkedin.com/in/nuno-andrade-964b2416

10+ Years in Cloud Architecture & High-Performance AI

Experienced Backend Developer and Cloud Architect with over a decade of expertise defining architectures, building, and migrating mission-critical systems to cloud and edge environments. Passionate about practical, production-grade AI agents and open-weights models (Gemma 4, Gemini, Google ADK, Node.js, C++).

Core Industry Sectors

High-Volume Transaction Systems
Energy Sector & Smart Grids
Gaming Industry Backends
Computer Vision & Multimodal AI

Connect & Resources

LinkedIn: linkedin.com/in/nuno-andrade-964b2416
Blog & Articles: https://nja.dev
GitHub: github.com/nuno-joao-andrade
Beyond Code: Openwater Swimmer 🏊‍♂️

🤝 Let's connect on LinkedIn: www.linkedin.com/in/nuno-andrade-964b2416

🔗 Open LinkedIn Profile

⚠️ The Founder Pitch Challenge & Workshop Purpose

Audio-Visual Cognitive Dissonance in Startup Pitches

Why do 80% of early-stage pitches lose an Investment Committee in the first 30 seconds? Because spoken claims contradict visual slide evidence.

🎙️ What the Founder Says Out Loud (Audio)

“Our capital efficiency is best-in-class — our burn multiple is a rock-solid 1.0x, our LTV-to-CAC is 4.8x, and we are tackling a $90 billion bottom-up TAM across 150,000 mid-market companies paying $3,000 a year!”

Sounds confident, fast-paced, and investor-ready.
Text-only transcripts miss tone, pacing, and visual tables.

📊 What Slide #2 Actually Shows (Visuals)

When Gemma 4 inspects the financial table on Slide #2 at 1,120 vision tokens:

Net New ARR: $1.2M vs. Monthly Burn: $350K/mo ($4.2M/yr) ➔ Actual Burn Multiple = 3.50x!
Bottom-Up TAM Math: 150,000 × $3,000 = $450M (200x inflation vs. $90B claim!).
8pt Footnote: Excludes $190K/mo AWS GPU cluster costs.

🎯 PURPOSE OF THIS WORKSHOP: This tool is just a guide to better pitch your project — this isn't the real result. Use it to catch discrepancies in your own deck before you step on stage!

💡 Why Local AI? Founders won't upload stealth financial decks to third-party cloud APIs!

🏗️ System Architecture • 100% On-Device Privacy

Gemma 4 (gemma4:e2b) + Local Ollama Pipeline

Zero cloud API keys, zero network latency spikes on conference Wi-Fi, and zero data retention.

1. Spoken Pitch (.WAV) 16kHz 16-bit Mono PCM

+

2. Deck Slides (.PNG) 280 vs 1,120 Vision Tokens

➔

3. Local Ollama Server 127.0.0.1:11434 • gemma4:e2b

⇄

4. 4 VC Math Tools Deterministic Node.js

➔

5. Zod JSON Verdict Score + Contradictions + Roast

🧠 Configurable Thinking

We enable think: true on Ollama's /api/chat endpoint so gemma4:e2b plans its cross-modal audit inside message.thinking before emitting JSON.

🖼️ Vision Budget Control

In shared/visionBudget.js, we use sharp to map 280 tokens (448×448) vs. 1,120 tokens (1120×1120) before base64 encoding.

🔒 Strict Zod Schemas

In shared/schemas.js, we convert Zod v4 schemas via z.toJSONSchema() and pass them to Ollama's format parameter.

⚡ Shared Engine Location: shared/{client,audioUtils,visionBudget,tools,schemas}.js

🛠️ Pre-Flight Setup • Node.js & Local Ollama

Setting Up Node.js & Ollama (gemma4:e2b)

All dependencies are modern, vulnerability-free ES Modules (0 vulnerabilities on npm audit).

1. Install Ollama & Pull gemma4:e2b

# Linux install:
curl -fsSL https://ollama.com/install.sh | sh

# macOS (Homebrew) / Windows (winget):
brew install --cask ollama
# or: winget install Ollama.Ollama

# Pull the multimodal Gemma 4 E2B model (~3.1 GB):
ollama pull gemma4:e2b

# Verify local Ollama server (127.0.0.1:11434):
ollama list

2. Install Workshop Packages & Run Tests

# Ensure Node.js v20+ or v22 LTS is active:
node -v && npm -v

# Install root dependencies (0 vulnerabilities):
npm install

# Run all 11 automated tests (audio, vision, tools, Ollama):
npm test

# Optional: Override model or Ollama host anytime:
export OLLAMA_HOST=http://127.0.0.1:11434
export OLLAMA_MODEL=gemma4:e2b

📋 Click to copy setup commands:

🗺️ 90-Minute Workshop Roadmap

5 Self-Contained Step Folders & 3 Pitch Fixtures

Every step is ready to run with cd step-0X-... && npm start. Pre-built test audio & slides live in fixtures/.

Time	Step Folder	Command	What You Will Experience & Learn
00–15m	step-01-setup/	`cd step-01-setup && npm start`	Verify Local Ollama (gemma4:e2b) + Run the 20s Live Pitch Roast on QuantumSynergy.ai.
15–35m	step-02-narrative/	`cd step-02-narrative && npm start`	Sprint 1: Native 16kHz WAV ingestion + 5-beat Narrative Arc scoring with <think> & Zod JSON.
35–55m	step-03-contradictions/	`cd step-03-contradictions && npm start`	Sprint 2: Cross-Modal Contradiction Detection + 280 vs. 1,120 Vision Token Budget A/B test.
55–75m	step-04-tools/	`cd step-04-tools && npm start`	Sprint 3: 4 Deterministic VC Tool Calls (audit_unit_economics, verify_tam_math, etc.) + IC Verdict.
75–90m	step-05-arena/	`cd step-05-arena && npm start`	The Live Pitch Arena Finale: Full Portugal Tech Week Web UI (Record 30s Mic / Upload WAV + Slides).

🔥 Fixture 1: QuantumSynergy.ai (20s) Zero churn claim vs 18.4% slide churn • 11 buzzwords ➔ HARD PASS

🧪 Fixture 2: CloudMetrics Pro (30s) 1.0x spoken burn vs 3.5x slide burn • 200x TAM inflation • 8pt footnote

✅ Fixture 3: VeraGrid Energy (30s) 0.75x burn multiple • 6.0x LTV:CAC • Consistent math ➔ INVEST

🚀 Ready for Step 1? Let's roast our first pitch!

🔥 Step 1 (00–15m) • step-01-setup/

The 20-Second Pitch Roast & Health Check

File: step-01-setup/index.js • Fixture: fixtures/fixture_01_roast/ (QuantumSynergy.ai)

step-01-setup/index.js

// 1. Verify Local Ollama (http://127.0.0.1:11434)
const health = await checkOllamaHealth();

// 2. Load 16kHz Mono PCM WAV + Slide 1 (1120 tokens)
const audio = loadAndNormalizeWav(fixture.audioPath);
const slide1 = await prepareSlideImage(
  fixture.slides[0].path,
  1120 // High-res vision budget
);

// 3. Run full multimodal + tool-calling audit
const { verdict } = await runFullPitchArenaPipeline({
  startupName: 'QuantumSynergy.ai',
  audioPath: fixture.audioPath,
  slides: [{ path: fixture.slides[0].path, budget: 1120 }]
});

🎤 Spoken Claim vs. Slide #1 Reality

Spoken @ 00:05: “Zero churn & 4,200 enterprise pilots!”
Slide #1 Table: 18.4% Monthly Logo Churn & 3 Unpaid Design Partners ($14.2K ARR).
Spoken @ 00:11: “Infinite LTV!” vs. Slide $48.5K CAC / $9.2K LTV (0.19x).

🔥 Resulting IC Verdict: HARD PASS (Score: 12/100)

Tools Triggered: flag_buzzword_overload catches 11 buzzwords in 19 seconds; audit_unit_economics flags a 164.8x annual burn multiple.

▶️ Run Step 1 in your terminal:

🎙️ Step 2 (15–35m) • Sprint 1 • step-02-narrative/

Native 16kHz Audio Ingestion & Narrative Arc Scoring

File: step-02-narrative/index.js & shared/audioUtils.js • Scoring the 5 Pitch Beats (0–10)

shared/audioUtils.js + step-02-narrative/index.js

// Normalize any .wav to 16kHz, 16-bit Mono PCM
const wav = new WaveFile(rawBuffer);
wav.toSampleRate(16000);
wav.toBitDepth('16');

// Call Local Ollama with thinking + Zod JSON Schema
const response = await ollamaChat({
  messages: [
    { role: 'system', content: NARRATIVE_SYSTEM_PROMPT },
    { role: 'user', content: prompt, audio: [audio.base64] }
  ],
  think: true,
  format: toOllamaJsonSchema(NarrativeStructureSchema)
});

📊 5 Scored Narrative Beats (shared/schemas.js)

1. Hook (0–10): Opening attention-grabber & timestamp range.
2. Problem (0–10): Customer pain clarity & urgency.
3. Solution (0–10): Product mechanism & differentiation.
4. Traction (0–10): ARR, growth & unit economics claims.
5. The Ask (0–10): Fundraising amount & 18-month milestones.

🧠 Why Enable think: true?

Gemma 4 segments the pitch timeline chronologically and evaluates vocal pacing (WPM) inside <think> before committing to numeric scores.

▶️ Run Sprint 1 (Narrative Structure Analyzer):

🖼️ Engineering Deep Dive • shared/visionBudget.js

Variable Vision Token Budgets: 280 vs. 1,120 Tokens

Why waste 4× compute on a minimalist title slide—or lose critical 8pt financial footnotes by downsampling a dense table?

⚡ 280 Vision Tokens (448×448 px • Layout Mode)

Best For: Title slides, hero value propositions, simple 3-bullet layouts (Slide #1).
Latency & KV Cache: ~4× faster prefill and minimal RAM footprint on laptop GPUs/CPUs.
Trade-off: Small 8pt fine-print footnotes blur below OCR legibility threshold!

🔬 1,120 Vision Tokens (1120×1120 px • Dense Mode)

Best For: Financial tables, cohort retention charts, dense architecture diagrams (Slide #2 & #3).
High-Res OCR: Reads every cell in CloudMetrics Pro Slide #2 plus the 8pt bottom footnote:
“*Footnote: Spoken pitch figures exclude $190k/mo AWS GPU cluster costs and 4.2% monthly logo churn.”

shared/visionBudget.js — Resizing Slide Images to Match Token Budgets

export const VISION_BUDGET_CONFIG = {
  280:  { tokens: 280,  targetDimension: 448,  label: 'LOW_280_LAYOUT' },
  1120: { tokens: 1120, targetDimension: 1120, label: 'HIGH_1120_DENSE_OCR' }
};

💡 In Step 3, we run an A/B comparison on Slide #2 at 280 vs 1,120 tokens!

⚡ Step 3 (35–55m) • Sprint 2 • step-03-contradictions/

Cross-Modal Contradiction Detection (Audio × Slides)

File: step-03-contradictions/index.js • Fixture: fixtures/fixture_02_sprint/ (CloudMetrics Pro)

step-03-contradictions/index.js

// Part A: Compare 280 vs 1120 Vision Tokens on Slide 2
const slide2Low  = await prepareSlideImage(slide2Path, 280);
const slide2High = await prepareSlideImage(slide2Path, 1120);

// Part B: Full Multimodal Contradiction Audit
const response = await ollamaChat({
  messages: [{
    role: 'user',
    content: buildContradictionPrompt(startupName, slideMeta),
    images: preparedSlides.map(s => s.base64),
    audio: [audio.base64]
  }],
  think: true,
  format: toOllamaJsonSchema(ContradictionAuditSchema)
});

🚨 Contradiction #1: Burn Multiple Lie (FATAL_BS)

Spoken @ 00:08: “Burn multiple is a rock-solid 1.0x”
Slide #2 Evidence: $350K/mo burn ($4.2M/yr) ÷ $1.2M Net New ARR = 3.50x burn multiple.

🚨 Contradiction #2: 200x TAM Inflation (FATAL_BS)

Spoken @ 00:18: “$90B bottom-up TAM”
Slide #2 Evidence: 150,000 customers × $3,000 ACV = $450M actual TAM.

🚨 Contradiction #3: LTV:CAC Exaggeration (MAJOR)

Spoken @ 00:13: “4.8x LTV:CAC” vs. Slide #2: $31,200 LTV ÷ $18,400 CAC = 1.70x.

▶️ Run Sprint 2 (Multimodal Contradictions & Vision Budget A/B Test):

🛠️ Engineering Deep Dive • shared/tools.js

The 4 Deterministic VC “Bullshit Detector” Tools

Let Gemma 4 extract the numbers from audio & slides — let deterministic Node.js functions execute the exact math!

1. `audit_unit_economics`

Computes (monthlyBurn × 12) / netNewArr and ltv / cac. Flags FAIL if Burn Multiple > 3.0x or deviates from spoken claim by > 0.25x.

2. `verify_tam_math`

Computes bottom-up TAM: customers × annualPrice. Flags FAIL (TAM_INFLATED) if claimed TAM exceeds bottom-up math by > 1.25x.

3. `check_slide_density`

Audits words & data points per slide. Flags FAIL (WALL_OF_TEXT) if > 75 words (e.g., Slide #3 with 96 words) and recommends 1120 tokens.

4. `flag_buzzword_overload`

Scans transcript & slides for empty VC buzzwords (quantum-ready, autonomous mesh, infinite LTV) and calculates hits/minute.

shared/tools.js — Deterministic Burn Multiple & LTV:CAC Verification

const annualNetBurnUsd = monthlyNetBurnUsd * 12;
const actualBurnMultiple = Number((annualNetBurnUsd / netNewArrUsd).toFixed(2));
const actualLtvCacRatio  = Number((ltvUsd / cacUsd).toFixed(2));
const isCritical = actualBurnMultiple > 3.0 || actualLtvCacRatio < 1.5 || burnDiscrepancy > 0.5;

🧪 All 4 tools are unit-tested in tests/tools.test.js

⚙️ Step 4 (55–75m) • Sprint 3 • step-04-tools/

Multi-Turn Ollama Tool Loop & Final IC Verdict

File: step-04-tools/index.js & shared/client.js • Grounding the Final Verdict in Deterministic Math

shared/client.js — 2-Turn Ollama Tool Loop

// Turn 1: Request tool calls from gemma4:e2b
const turn1 = await ollamaChat({
  messages: toolMessages,
  tools: OLLAMA_VC_TOOLS,
  think: true
});

// Execute each requested tool deterministically in Node.js
for (const call of turn1.message.tool_calls) {
  const result = executeVcTool(call.function.name, call.function.arguments);
  toolMessages.push({
    role: 'tool',
    tool_name: call.function.name,
    content: JSON.stringify(result)
  });
}

// Turn 2: Synthesize final Zod-validated IC Verdict

📋 Tool Execution Results on CloudMetrics Pro

audit_unit_economics ➔ FAIL (Actual Burn Multiple 3.50x vs 1.0x claimed; LTV:CAC 1.70x vs 4.8x).
verify_tam_math ➔ FAIL ($90B claimed vs $450M bottom-up = 200x inflation).
check_slide_density ➔ FAIL (Slide #3 has 96 words — Wall of Text).
flag_buzzword_overload ➔ PASS (Only 2 buzzwords).

🎯 Final Grounded Output (InvestmentCommitteeVerdictSchema)

Combines verdict (HARD_PASS / DIG_DEEPER / INVEST), overallIcScore, liveRoast, and topFixRecommendation.

▶️ Run Sprint 3 (Deterministic VC Tools & Final Verdict):

🏆 Step 5 (75–90m) • Finale • step-05-arena/

The Live Portugal Tech Week Pitch Arena UI

Full-stack Express + React 19 + Vite + Tailwind CSS v4 Web App at http://localhost:5173 (API at :4000).

1. Record or Upload Audio

Use ● Record 30s Pitch (Mic) (browser Web Audio resampled in-memory to 16kHz mono WAV) or click ↑ Upload Audio (.wav), or load a 1-click PTW Demo Preset.

2. Upload Slides & Set Budgets

Drag & drop up to 3 slide images (PNG/JPG/WEBP), edit slide titles, and toggle each slide between 280 Tokens (Layout) and 1,120 Tokens (Dense).

3. Live SSE Streaming Verdict

Watch real-time Server-Sent Events (POST /api/pitch/analyze-stream) stream Gemma 4's <think> trace, tool badges, contradictions, and Live Roast!

ℹ️ BUILT-IN UI DISCLAIMER DIALOG: The web app includes the Workshop Disclaimer Dialog reminding everyone that this is a workshop guide to better pitch your project, not a real competition result.

🚀 Launch the Full Portugal Tech Week Pitch Arena Web App:

🌐 Open http://localhost:5173

🇵🇹 Obrigado! • Portugal Tech Week 2026 Wrap-Up

Workshop Cheat Sheet & Stay Connected

You now have a complete, 100% local multimodal AI pitch coach running on your own machine with gemma4:e2b and Ollama.

Complete Workshop Command Cheat Sheet

# Run automated test suite (11/11 tests):
npm test

# Run each workshop step independently:
cd step-01-setup && npm start          # 20s Roast
cd step-02-narrative && npm start      # Audio Narrative Arc
cd step-03-contradictions && npm start # 280 vs 1120 Tokens
cd step-04-tools && npm start          # 4 Deterministic VC Tools
cd step-05-arena && npm start          # Full PTW Web Arena UI

# Open these workshop slides anytime:
npm run slides

👤 Connect with Nuno João Andrade

Nuno João Andrade

Senior Dev & Cloud Expert • Google Developer Expert (GDE)

🔗 www.linkedin.com/in/nuno-andrade-964b2416

🌐 https://nja.dev

🎯 Key Engineering Takeaways

Native Audio + Vision: gemma4:e2b processes 16kHz WAV & PNG slides locally in one pass.
Right-Size Vision Tokens: Use 280 for layout slides and 1,120 for dense tables/footnotes.
Deterministic Tool Calling: Delegate financial ratios & TAM math to verified code tools.

🇵🇹 Thank you, Portugal Tech Week! Connect on LinkedIn:

🔗 www.linkedin.com/in/nuno-andrade-964b2416

Build a Multimodal Pitch & Deck Critic with Gemma 4 & Local Ollama

Native Audio Ingestion

Variable Vision Budgets

Deterministic VC Tools

Nuno João Andrade

10+ Years in Cloud Architecture & High-Performance AI

Core Industry Sectors

Connect & Resources

Audio-Visual Cognitive Dissonance in Startup Pitches

🎙️ What the Founder Says Out Loud (Audio)

📊 What Slide #2 Actually Shows (Visuals)

Gemma 4 (gemma4:e2b) + Local Ollama Pipeline

🧠 Configurable Thinking

🖼️ Vision Budget Control

🔒 Strict Zod Schemas

Setting Up Node.js & Ollama (gemma4:e2b)

5 Self-Contained Step Folders & 3 Pitch Fixtures

The 20-Second Pitch Roast & Health Check

🎤 Spoken Claim vs. Slide #1 Reality

🔥 Resulting IC Verdict: HARD PASS (Score: 12/100)

Native 16kHz Audio Ingestion & Narrative Arc Scoring

📊 5 Scored Narrative Beats (shared/schemas.js)

🧠 Why Enable think: true?

Variable Vision Token Budgets: 280 vs. 1,120 Tokens

⚡ 280 Vision Tokens (448×448 px • Layout Mode)

🔬 1,120 Vision Tokens (1120×1120 px • Dense Mode)

Cross-Modal Contradiction Detection (Audio × Slides)

🚨 Contradiction #1: Burn Multiple Lie (FATAL_BS)

🚨 Contradiction #2: 200x TAM Inflation (FATAL_BS)

🚨 Contradiction #3: LTV:CAC Exaggeration (MAJOR)

The 4 Deterministic VC “Bullshit Detector” Tools

1. audit_unit_economics

2. verify_tam_math

3. check_slide_density

4. flag_buzzword_overload

Multi-Turn Ollama Tool Loop & Final IC Verdict

📋 Tool Execution Results on CloudMetrics Pro

🎯 Final Grounded Output (InvestmentCommitteeVerdictSchema)

The Live Portugal Tech Week Pitch Arena UI

1. Record or Upload Audio

2. Upload Slides & Set Budgets

3. Live SSE Streaming Verdict

Workshop Cheat Sheet & Stay Connected

👤 Connect with Nuno João Andrade

🎯 Key Engineering Takeaways

Build a Multimodal Pitch & Deck Critic
with Gemma 4 & Local Ollama

1. `audit_unit_economics`

2. `verify_tam_math`

3. `check_slide_density`

4. `flag_buzzword_overload`