Research Navigator — Vali

Research Navigator's Vali is an autonomous co-pilot powered by Gemini 2.0 Flash & Google ADK that turns static market ideas into proven Caribbean startup evidence in real time.

Comment

💡 Inspiration

Most early-stage startups don’t fail because they can't write code—they fail because they build things nobody actually wants.

Founders conduct customer discovery interviews, but fall into the "Polite Lie" trap: they ask hypothetical, leading questions ("Would you buy this if we built it?") and hear flattering, non-binding compliments. Furthermore, in emerging markets like the Caribbean, founders don't carry laptops to the market or store meetings; they capture thoughts via WhatsApp voice notes on the go, leaving recordings unanalyzed.

We built Vali to act as a relentless, active Collaborative Partner for founders. Vali doesn't just generate static reports—it coaches founders in real time during customer interviews, enforces The Mom Test, transcribes multimodal voice notes from the field, and calculates mathematically grounded Market Conviction Scores.

🚀 What It Does

Vali operates collaboratively across two human-agent touchpoints:

  1. 🖥️ In-Canvas Live Mom Test Coach (Web Platform):

    • As founders draft questions in the Interview Canvas, Vali detects hypothetical or leading questions in <300ms using Gemini 2.0 Flash.
    • It issues real-time warnings and suggests 1-click drill-down probes grounded in past behavior, workarounds, and actual dollars spent.

  2. 📱 Field Co-Pilot over WhatsApp (Multimodal Audio & RAG):

    • Founders send voice notes directly after customer meetings. Vali uses native Gemini 2.0 Flash Multimodal Audio understanding to transcribe regional dialects accurately and extract concrete numbers, costs, and quotes into the project's evidence buffer.

  3. 🔍 Grounded Lead Finder (Google Search Tool Grounding):

    • Discovers real Caribbean retail, commercial, and enterprise contacts using Google Search Tool Grounding (types.GoogleSearch()) with tailored, non-pitch discovery outreach templates.

  4. 📊 Multi-Interview Discovery Synthesis:

    • Aggregates qualitative interviews across a validation task, outputting an objective Market Conviction Score (0-100%), validated pain clusters, existing workaround spend, and roadmap gating recommendations.

🛠️ How We Built It & Architectural Evolution

Vali represents an intentional shift from static AI generation to dynamic human-agent collaboration:

  • Evolving from a Sequential Pipeline to Standalone ADK Agents: Prior to the hackathon, Research Navigator relied on a linear, sequential research pipeline that executed static waterfall steps. During this hackathon, we refactored and decomposed that sequential pipeline into dedicated, standalone autonomous agents orchestrated dynamically by the Google Agent Development Kit (ADK) and Google GenAI SDK.
  • Google Gemini 2.0 Flash: Primary reasoning and multimodal engine powering sub-second live coaching, audio transcription, and multi-interview synthesis.
  • Google GenAI SDK (google-genai): Native Python client using structured output schemas and type-safe generation configs.
  • Google Search Tool Grounding: Integrated types.Tool(google_search=types.GoogleSearch()) for hallucination-free business lead discovery.
  • Google Agent Development Kit (ADK): Orchestrates standalone agents with multi-turn stateful conversational loops, memory retention, and proactive check-in patterns.
  • Google Cloud Run: Serverless container deployment (us-central1) with scale-to-zero efficiency configured via cloudbuild.yaml.
  • Full-Stack Orchestrator: Go backend orchestrator (api-go), Vue 3 authenticated frontend (platform), and Meta WhatsApp Cloud API webhooks.

📦 Third-Party & Pre-Existing Code Disclosure

  • Pre-Existing Code: The Research Navigator baseline web platform (Vue 3 shell and Go backend orchestrator) existed as our foundation, along with an initial sequential research pipeline.
  • Built During Hackathon: During the hackathon, we decomposed the sequential pipeline into standalone autonomous agents orchestrated by Google ADK, built the entire Vali Autonomous Co-Pilot Agent (services/vali-agent), Live In-Canvas Mom Test Coach (ValiLiveCoachDrawer.vue), Multi-Interview Discovery Synthesizer (DiscoverySynthesisModal.vue), Multimodal WhatsApp Audio Processing pipeline, and Google Search Grounding integration from scratch.

🧗 Challenges We Ran Into

  • Dialect & Dialectical Preservation: Transcribing Caribbean English and business terms without standardizing away critical nuance. Gemini 2.0 Flash's multimodal audio understanding preserved local business terminology flawlessly.
  • Real-Time Coaching Latency: Founders typing in an interview canvas cannot wait 3–5 seconds for coaching. By leveraging Gemini 2.0 Flash with targeted system instructions and structured JSON output, we achieved sub-second (<400ms) leading question detection.
  • Synthesizing Unstructured Evidence: Translating qualitative conversation quotes into an objective, trustworthy conviction score. We developed a multi-stage aggregation pipeline that extracts verified past actions, clusters pain frequencies, and computes an evidence-backed score.

🏆 Accomplishments That We're Proud Of

  • True Omnichannel Collaboration: Seamless handoff between field voice notes on WhatsApp and desktop synthesis in the web canvas.
  • Live Mom Test Guardrails: An in-browser co-pilot drawer that actively protects founders from cognitive bias and false positive feedback.
  • Grounded Regional Discovery: Utilizing Google Search Grounding to find actual business contacts in Trinidad, Jamaica, and the wider Caribbean.

📚 What We Learned

We learned that the future of startup tooling isn't static "generate me a business plan" generators—it's collaborative, conversational partners that stay with founders turn-by-turn, challenge their assumptions, and ensure they stay honest to customer reality.

🔮 What's Next for Vali

  • Live Meeting Audio Coaching: Real-time audio analysis during live Google Meet discovery calls.
  • Automated Cold Outreach: Generating personalized, non-pitch outreach emails grounded in verified customer pain points.
  • Ecosystem Sync: 1-click export of validated customer evidence into Linear, Notion, and Jira backlogs.

Built With

Share this project:

Updates