Proxima

Multimodal AI prospects for realistic sales rehearsal

Comment 1

Proxima

Proxima turns static role-play into a live, adaptive rehearsal environment where BDRs can safely make mistakes, sharpen their instincts, and walk into real calls already “warmed up.”

Inspiration

Most sales coaching happens after the fact—through recordings, scorecards, and retroactive feedback—when the opportunity to adjust is already gone.

Proxima shifts training before the call, creating a space where reps can rehearse full conversations that feel realistic and context-driven.

Advances in real-time generative AI, multimodal reasoning, and voice models make it possible to simulate:

  • Natural, interruptible conversations
  • Context-aware prospect behavior
  • Tool-augmented reasoning grounded in real deal data

What Proxima Does

Proxima is a real-time, AI-driven sales rehearsal platform for high-velocity revenue teams.

It enables reps to prepare, practice, and perform through:

  • Lifelike sales call simulations using context-rich personas
  • Real-time, bidirectional voice conversations
  • Interruption handling and dynamic topic control
  • Live coaching overlays during conversations
  • Post-session analytics combining quantitative and qualitative insights
  • Multi-participant AI simulations with teammates and stakeholders
  • Multimodal grounding using files, images, and structured inputs

Core Screens

Dashboard

  • View upcoming rehearsals
  • Access recent sessions
  • Review coaching reports
  • Track performance trends over time

Dashboard

Context Builder

Structured environment for defining simulation inputs and shaping AI behavior.

Context Builder

Persona Definition

  • Prospect identity (job title, company, industry, location)
  • Funnel stage (Awareness, Consideration, Decision, Expansion)
  • Stage intent (specific goal of the conversation)

Behavioral Modeling

  • Objection archetypes (Skeptic, Visionary, Guardian)
  • Decision styles (Data-driven, Consensus-based, Intuitive, Top-down)
  • Tunable sliders:
    • Skepticism level
    • Negotiation toughness
    • Initial trust score

Context Injection

  • Key-value inputs for attaching assets such as:
    • Decks
    • RFP summaries
    • Pricing sheets

Persona Randomization

  • Automatically generates realistic variability across:
    • Stage
    • Archetype
    • Decision style
    • Behavioral sliders

AI Teammates

  • Toggle participation on/off
  • Assign roles or randomize behavior
  • Simulate internal dynamics (e.g., dominant AE, hesitant junior rep)
    Multi-participant training

Live Session Room

Primary simulation environment for real-time interaction.

  • Video-style layout with:

    • User
    • AI prospect
    • AI teammates

  • Live transcript and coaching sidebar

  • Real-time feedback signals:

    • Missed discovery opportunities
    • Weak responses
    • Over-talking or lack of clarity

  • Controls for:

    • Mute
    • Screen sharing
    • Session termination

Live Session

Coaching

Session Report

Structured post-session analysis.

  • Talk ratio and listening balance
  • Question quality and discovery depth
  • Objection handling effectiveness
  • Coverage of key conversation areas

Includes:

  • Timestamped transcript highlights
  • Actionable improvement suggestions
  • Reusable coaching insights

Session Report

Persona Library

Reusable catalog of prospect profiles. Enables standardized training scenarios

Persona Library

Sessions List

Historical view of all rehearsals.

  • Chronological session tracking
  • Quick access to:
    • Replays
    • Reports
    • Performance trends

Sessions List

How Everything Connects

Proxima is built as a tightly integrated real-time system powered by Google Cloud Platform and Gemini models.

  • The Next.js client streams audio, transcripts, and events via WebSockets
  • The FastAPI backend manages session state and orchestration
  • AI models handle conversation, reasoning, and analysis
  • Firestore persists structured data across sessions

This creates a continuous loop:

Configure → Simulate → Coach → Analyze → Improve

Tech Stack and Architecture

Frontend

  • Next.js (React) application
  • WebSocket-based streaming for low-latency interaction

Backend

  • FastAPI (Python) services for:

    • Session orchestration
    • Persona generation
    • Report generation

  • Integrated with Google Gen AI Python SDK

AI and Models (Gemini-Centric)

  • Gemini Live (via Vertex AI)

    • Real-time, bidirectional voice conversations
    • Handles interruption, turn-taking, and tool invocation

  • Gemini Multimodal

    • Processes text, images, and files
    • Powers persona grounding and report generation

  • Tooling Layer

    • Coaching hint generation
    • File summarization
    • Knowledge retrieval
    • UI event triggering

All model interactions are orchestrated through Google’s Gen AI SDK, ensuring tight integration with GCP services.

GCP Infrastructure (Core Highlight)

  • Vertex AI

    • Managed runtime for Gemini models
    • Scalable inference for both live and batch workloads

  • Cloud Run

    • Containerized backend deployment
    • Horizontal scaling for concurrent sessions

  • Firestore

    • Source of truth for:
    • Personas
    • Sessions
    • Reports
    • Configuration

This architecture ensures:

  • Low latency
  • High concurrency
  • Production-ready scalability

Integration Layer

  • WebSockets for real-time streaming
  • REST APIs for structured operations
  • Extensible tool registry for adding new capabilities

Sessions List

Challenges

Real-time AI Orchestration

  • Managing simultaneous voice, text, and tool outputs required a suppression and arbitration layer

Multi-Participant Simulation

  • Coordinating multiple AI agents with shared context and overlapping audio streams

Interruption Handling

  • Supporting natural barge-in without breaking conversational coherence

Persona Consistency

  • Maintaining behavioral realism across turns and sessions

Hybrid Analytics

  • Combining deterministic metrics with qualitative AI evaluation

Multimodal Context Flow

  • Aligning file uploads, screen shares, and live context in real time

Accomplishments

  • Delivered low-latency, real-time conversational performance
  • Built a dynamic persona engine grounded in real inputs
  • Implemented live coaching without disrupting flow
  • Enabled multi-agent simulation environments
  • Integrated multimodal reasoning for richer context understanding

Roadmap

Playbook Integration

  • Embed company-specific sales playbooks into simulations
  • Reinforce real-world selling frameworks

Knowledge Graph Reasoning

  • Model relationships between:

    • Prospects
    • Companies
    • Opportunities

  • Improve contextual awareness and follow-up quality

Competitive Intelligence

  • Simulate real-world objections and competitor narratives
  • Dynamically scale difficulty

Role Switching

  • Allow reps to play the prospect
  • Identify weaknesses in messaging

Advanced Analytics

  • Funnel-stage heatmaps
  • Individual and team performance tracking
  • Correlation with pipeline outcomes

Localization

  • Multi-language simulations
  • Region-specific tone and objection styles

Emotion Modeling

  • Expressive AI avatars with behavioral cues
  • Train reps on timing and emotional intelligence

Live Meeting Integration

  • Join real calls as an AI assistant
  • Provide real-time coaching and post-call analysis

Adaptive Training (Long-Term Vision)

  • Multi-session, branching simulations
  • Dynamic difficulty adjustment
  • Automated playbook optimization from aggregated data

Built With

Share this project:

Updates