Spicedream

Inspiration

We noticed that our co-founder, Adam, was spending 20–30 hours a week on inbound sales calls with people who often weren’t qualified or clear on what they wanted. We thought, “What if an AI could handle these repetitive interactions, qualify leads, and capture key information automatically?” That spark of an idea—to build an AI version of Adam—drove us to create Spicedream.

What it does

Spicedream is an AI-driven sales-call simulator that fields inbound inquiries and answers typical questions about our AI consulting services. It: 1. Simulates a live conversation with “AI Adam” (powered by Tavis + 11Labs). 2. Gathers important lead information—automatically populating our CRM so we can see who’s qualified and worth following up with. 3. Handles booking and scheduling through levelable.dev, meaning our “AI Adam” can meet users anytime—no human availability required. 4. Provides LLM observability via AgentOps to monitor performance and track conversation data effectively.

How we built it 1. Tavis + 11Labs for AI Voice • Tavis powers the real-time AI interactions, while 11Labs handles the realistic voice generation, giving callers an experience that feels personal and human. 2. levelable.dev for Calendar Booking • Users can schedule a “call” with AI Adam any time of day—no calendar conflicts because no real person is needed. 3. AgentOps for Observability • We implemented AgentOps to better analyze how AI calls flow, spot potential conversation bottlenecks, and improve the overall user journey. It provided deeper insights than PostHog for our use case. 4. Prompt Engineering • We spent hours refining prompts to ensure the AI’s tone is polite, professional, and self-aware. We also used “SPICE” methodology to guide conversation structure and qualification steps.

Challenges we ran into • Tavis API Limitations: Calls currently cap at 5 minutes (max five questions) due to an apparent bug. We couldn’t extend the duration, which impacted demos. • Prompt Tuning: Achieving the right balance between friendly conversation and strict qualification criteria took significant prompt engineering. We needed the AI to appear both approachable and authoritative in AI consulting. • Integrations: Syncing real-time conversations with our CRM and ensuring calendar logic worked smoothly introduced more complexity than expected.

Accomplishments that we’re proud of • Fully Automated Sales Calls: We’ve drastically cut down on Adam’s inbound call time, freeing him to work on high-value tasks. • Accurate Lead Qualification: By capturing relevant data, we can quickly identify and follow up with strong leads. • Scalable Scheduling: levelable.dev’s integration means clients can “meet” AI Adam 24/7 without ever clashing with an actual human schedule. • Improved Observability: Thanks to AgentOps, we have a clear view of conversation flow and performance metrics, helping us refine the user experience.

What we learned • Prompt Engineering is Key: Even small prompt tweaks had outsized effects on how the AI responded and steered conversations. • Observability Matters: LLM-driven products benefit hugely from robust data tracking—AgentOps’ in-depth analytics gave us actionable insights. • API Limitations can be Deal-Breakers: Building an AI product means sometimes working around third-party constraints or bugs. Choosing partners and tools with flexibility and good support is crucial.

What’s next for Spicedream • Extend Call Length: We’ll continue engaging with Tavis to fix or work around the five-minute limit. • Refine SPICE: Further iteration on our SPICE methodology to create even more targeted and streamlined prompts. • Advanced CRM Integration: Deeper insights by automatically tagging leads, generating qualification scores, and triggering follow-up sequences. • Multilingual Support: Exploring voice support in multiple languages to expand the reach of AI Adam and accommodate a global user base.

Built With

• elevenlabs
• lovable
• python
• tavus