SLEM

• 

Inspiration

What it does

How we built it

Challenges we ran into

Inspiration We are in a "Soft Skills Recession." While digital fluency is skyrocketing, interpersonal confidence is plummeting. Sales reps and support agents are being trained with passive videos and multiple-choice quizzes, which is like trying to learn boxing by reading a book.

We realized that the only way to actually learn negotiation is through stress inoculation. We wanted to build a "Flight Simulator" for difficult conversations—a safe place to crash and burn so you don't do it on a real client call. We didn't want a polite chatbot; we wanted a simulator that punches back.

What it does SLEM (Say Less, Earn More) is a voice-first negotiation dojo.

The Sparring Partner: Users are connected via low-latency audio to "Linda," a hyper-realistic, furious CFO AI (powered by ElevenLabs).

The Stress Test: The AI is engineered to be difficult. It interrupts, demands specific data, and uses "Strike" logic. If you are vague, rude, or slow, Linda hangs up on you.

The Logic Gates: Unlike standard chatbots, our agent follows strict behavioral protocols. It won't accept a refund until you use specific past-tense triggers ("I have sent it"), forcing the user to learn precise communication.

The Scorecard: Post-call, the system analyzes the transcript against proven negotiation frameworks (Validation, Calibrated Questions, Defensiveness) and assigns a strict 1-10 performance score.

How we built it We prioritized speed and immersion over complex middleware.

Frontend: Built with React and Tailwind (via Lovable) for a premium, "Glassmorphism" aesthetic that feels like high-end enterprise software.

Voice Engine: We utilized ElevenLabs Conversational AI for the full stack (Speech-to-Text, LLM, Text-to-Speech). We chose the Turbo v2.5 model for sub-second latency.

Backend Logic: We used Supabase not just for data, but as the logic layer. We implemented custom RPC functions to handle secure token generation and protect our API keys, bypassing the need for a heavy Node.js server.

Prompt Engineering: This was the hardest part. We implemented a "State Machine" within the System Prompt, using explicit phases (The Fight -> The Parking Lot -> The Victory) to prevent the LLM from hallucinating solutions.

Challenges we ran into The "Nice AI" Problem: LLMs are naturally sycophantic. Making the AI genuinely "mean" and willing to hang up on the user required fighting the model's safety training with strict "Tool Prohibition" rules.

The "Zombie" Hangup: We struggled with a race condition where the AI would trigger the "End Call" tool but keep speaking, causing the audio to cut off mid-sentence. We solved this by implementing a "Semantic Firewall" in the frontend that verifies the AI has actually said "Goodbye" before killing the connection, and adding a smart 6-10 second buffer to the teardown sequence.

Hallucinated Victory: The AI kept assuming the refund was processed just because the user said "I will do it." We had to program strict tense-detection (Future vs. Past) into the prompt to force the AI to reject vague promises.

Accomplishments that we're proud of The "Liar Paradox" Solution: We successfully engineered a guardrail that prevents the AI from triggering the hang-up tool when it says phrases like "I am not hanging up."

Latency: We achieved a conversational feel that allows for interruptions and "barge-in," making the argument feel visceral.

Visual Feedback: The "Breathing Orb" visualizer isn't just a loop; it's a reactive canvas that changes state based on who is holding the floor, solving the "Turn-Taking" UX problem common in voice apps.

What we learned Prompt Engineering is Coding: You cannot just ask an LLM to "be angry." You have to define variables, logic gates, and failure states within the text prompt itself.

Silence is Loud: Handling the silence while the user looks up data (simulated) was crucial. We had to teach the AI to wait aggressively ("I'm waiting...") rather than filling the void with small talk.

What's next for SLEM Enterprise Customization: Allowing companies to upload their own angry emails/tickets to generate custom "Bosses."

Biometric Sync: Integrating with Apple Watch to measure the user's heart rate during the argument. If they panic, the score drops.

God Mode: A manager view that can listen in live and whisper suggestions into the trainee's ear.

Accomplishments that we're proud of

What we learned

What's next for SLEM

Built With

• elevenlabs
• gemini
• lovable
• supabase