Sum-Up

Inspiration

Sum-Up started when a very simple question was asked: what is the biggest time sink for sales representatives? Both Teo and Frank had formerly worked in sales, and as we discussed, the answer seemed clear—it was the creation of follow-up emails after meetings. A necessity, but nonetheless tedious and a massive time sink.

This is why Sum-Up was created: a B2C software solution that ensures two things. First, that sales reps can be far more effective by automating email writing for them. Second, it ensures accuracy by summarizing key points and guaranteeing that even the most minute detail isn't overlooked. This product gets salespeople back to what they do best—selling their product in a personable way.

What it does

Sum-Up transforms sales call transcripts into actionable insights in seconds. Upload a .txt transcript of your sales call, and Sum-Up delivers:

Structured JSON Output - A comprehensive breakdown containing:
- Call summary with overview and participants
- Key discussion points (pain points, budget, timeline)
- Action items with clear ownership and deadlines
- Next steps to move the deal forward
Personalized Follow-Up Email - A natural, human-sounding email that:
- Thanks the prospect for their time
- References specific moments from the conversation
- Clearly outlines next steps
- Maintains a warm, professional tone (not robotic!)

What used to take 15-20 minutes of manual work now happens in under 30 seconds. No more forgetting action items, no more generic templates, no more time wasted on administrative tasks.

How we built it

Tech Stack:

Frontend: React (Create React App) with Tailwind CSS for a modern, responsive UI
AI Engine: NeuralSeek's enterprise RAG (Retrieval Augmented Generation) platform
File Processing: React-dropzone for seamless file upload, FileReader API for transcript parsing
API Integration: Axios for reliable HTTP communication with NeuralSeek
Icons & UX: Lucide-react for polished visual elements

Architecture:

Service Layer - Built a custom neuralseekService.js that orchestrates two distinct AI operations:
- Extraction Pipeline (Temperature: 0.3) - Structured data extraction for consistent JSON output
- Generation Pipeline (Temperature: 0.7) - Creative email writing for natural language
Component Structure - Modular React components for scalability:
- FileUpload: Drag-and-drop transcript ingestion
- ProcessingStatus: Real-time feedback during AI processing
- ResultsDisplay: Dual-pane output (JSON + Email) with copy/download functionality
Knowledge Base - Created comprehensive sales context documentation including:
- Action item identification patterns
- Email tone guidelines (conversational vs. formal)
- Common sales call patterns (discovery, demo, negotiation, check-in)
- Pain point language recognition

Development Environment:

Built on Ubuntu WSL 2 for optimal Node.js performance
VS Code with Remote-WSL extension for seamless development
Environment variables for secure API key management

Key Implementation Decisions:

Used lower temperature (0.3) for structured extraction to ensure consistency
Used higher temperature (0.7) for email generation to maintain natural, human tone
Implemented response cleaning to handle markdown code blocks from AI responses
Added comprehensive error handling with user-friendly feedback
Designed mobile-first responsive interface

Challenges we ran into

1. WSL Development Environment Issues Getting Node.js and npm working properly on Ubuntu WSL proved challenging initially. We encountered directory access errors and WSL 1 compatibility issues. We solved this by:

Upgrading to WSL 2 for better filesystem performance
Installing Node.js via nvm (Node Version Manager) for cleaner dependency management
Ensuring all development happened in the WSL filesystem, not Windows directories

2. Prompt Engineering for Dual Outputs Balancing structured JSON extraction with creative email generation was tricky. Using the same temperature for both resulted in either:

Overly rigid emails that sounded robotic (low temperature)
Inconsistent JSON structure (high temperature)

We solved this by implementing a two-stage pipeline with distinct temperature settings for each task.

3. NeuralSeek Response Parsing The AI sometimes wrapped JSON responses in markdown code blocks (```json), which broke our parsing. We implemented robust cleaning logic to strip markdown formatting while preserving the actual JSON content.

4. Maintaining Natural Email Tone Early iterations produced emails that sounded too formal or template-like. We refined our knowledge base with explicit guidelines:

Use contractions (I'll, we'll, you're)
Reference specific conversation moments
Avoid corporate jargon
Write in first person

This dramatically improved the output quality and made emails sound genuinely human.

5. File Upload UX Implementing a smooth drag-and-drop experience that worked across browsers required careful attention to the FileReader API and react-dropzone configuration. We also had to handle edge cases like:

Files opened from Windows filesystem while in WSL
Large transcript files
Non-.txt file types

6. State Management Flow Coordinating the UI flow from upload → processing → results → reset required careful state management. We had to ensure:

Loading states displayed at the right times
Error messages cleared appropriately
Users could process multiple transcripts without refresh

Accomplishments that we're proud of

1. Real Time-Savings Impact We validated that Sum-Up reduces post-call work from 15-20 minutes to under 30 seconds—a 97% time reduction. For a sales rep with 5 calls per day, that's 1.5 hours saved daily, or 37.5 hours per month. That's nearly a full work week every month given back to selling!

2. Enterprise-Grade AI Implementation Successfully integrated NeuralSeek's RAG platform with a custom dual-pipeline architecture. The separation of extraction (structured) and generation (creative) tasks demonstrates sophisticated AI orchestration.

3. Polished User Experience Built a production-ready interface with:

Smooth drag-and-drop file upload
Real-time processing feedback
One-click copy/download functionality
Responsive design that works on any device
Professional gradient design with modern aesthetics

4. Natural Language Quality Our emails consistently sound human-written, not AI-generated. Judges and test users couldn't tell the emails were automated—exactly what we were aiming for.

5. Comprehensive Knowledge Base Created detailed sales context documentation that guides the AI to understand:

Sales call patterns and structures
Action item identification across different call types
Natural conversation flow and tone
Industry-specific terminology

6. Complete End-to-End Solution From file upload to deliverable outputs in a single, seamless workflow. No manual intervention required at any step.

7. Hackathon Development Speed Built a complete, functional product in under 24 hours, including:

Full React application
NeuralSeek integration
Multiple reusable components
Comprehensive error handling
Production-ready UI

What we learned

Technical Learnings:

1. RAG Architecture Benefits Working with NeuralSeek taught us the power of Retrieval Augmented Generation. By providing a curated knowledge base of sales context, the AI produces far more relevant outputs than generic LLMs. The combination of:

Domain-specific knowledge
Real-time retrieval
Large language model processing

creates outputs that are both accurate and contextually appropriate.

2. Temperature as a Precision Tool We learned that temperature isn't just a "creativity slider"—it's a precision instrument. For structured outputs (JSON), lower temperatures (0.3) ensure consistency. For creative outputs (emails), higher temperatures (0.7) maintain natural variation while still being professional.

3. Prompt Engineering is Critical The quality of AI output is directly proportional to prompt quality. We learned to:

Be extremely explicit about output format
Provide clear examples of desired tone
Use negative examples ("Don't do X")
Request specific JSON structures
Guide the AI away from common pitfalls

Business & Product Learnings:

1. Domain Expertise Matters Teo and Frank's sales experience was invaluable. Understanding the actual pain points—not assumed ones—guided every design decision. Real-world experience beats speculation every time.

2. Time Savings Compound Small efficiency gains per task create massive productivity improvements at scale. 15 minutes saved per call × 5 calls per day × 250 work days = 312 hours per year. That's nearly 8 full work weeks!

3. AI Should Augment, Not Replace Sum-Up doesn't try to be the salesperson—it handles the administrative burden so humans can focus on relationship building. The best AI products enhance human capabilities rather than attempting to replace them.

Team & Process Learnings:

1. Move Fast, Iterate Faster We built a mock service first to test the UI before connecting to NeuralSeek. This parallel development approach saved hours and prevented bottlenecks.

2. Documentation Accelerates Development Creating comprehensive component specifications upfront made implementation faster and reduced bugs. Clear requirements = clear code.

3. Focus on the Demo In hackathons, the demo is the product. We optimized for demonstrability:

Fast processing times
Visual polish
Clear value proposition
Memorable wow moments

What's next for Sum-Up

1. Multi-Format Support

Add support for audio transcription (MP3, WAV, M4A)
Integrate with Zoom, Google Meet, Microsoft Teams for automatic recording capture
Support video file upload with automatic transcription

2. CRM Integrations