Center for Employment Opportunities+Atherial
Proposal

CoachLink

Voice AI Coaching for Reentry

December 18, 2025

Section 1

About Us

Mike Fishbein

MIKE FISHBEIN

PARTNER

Eugene Leychenko

EUGENE LEYCHENKO

PARTNER

What We Do

Atherial builds custom AI applications and agents. We have a strong portfolio of voice AI projects including for training and coaching.

Relevant Portfolio

We have built voice AI sales coaching applications that simulate realistic prospect calls and provide actionable coaching. Sellers choose scenarios and skills to practice.

Demo: AI Sales Role-Play Training

Section 2

Project Overview

The Problems We're Solving

Limited Coaching Access

Incarcerated individuals rarely have access to high-quality coaching, especially for reentry preparation, employment readiness, and parole hearing prep.

No Private Practice Space

People can't practice interviews or difficult conversations privately before high-stakes real situations (parole hearings, job interviews).

Inconsistent Programming

Reentry programming is constrained by staffing and funding limitations, leading to inconsistent access.

High Stakes, One Chance

Many critical moments (parole hearings, first job interviews) are one-time opportunities with life-changing consequences.

The Solution: AI-powered coaching accessible through standard prison phone systems, available 24/7, providing unlimited practice opportunities.

What We Learned From Your Pilot

Your successful pilot with recently released individuals demonstrated:

  • ~3,000 calls over six months with ~500 participants
  • Phone-based access removes all barriers (no digital literacy, smartphones, or internet required)
  • VAPI.ai + Twilio architecture that works
  • Proven coaching content across three curriculum phases: finding a job, job interview process, and first 90 days of employment
  • Verbal navigation allows users to jump between modules naturally

Key insight: The coaching methodology already exists and works. We're adapting this proven system for correctional environments.

The MVP Scope

Scope
  • 1-3 pilot facilities
  • 25 concurrent calls
  • English + Spanish support
  • 8 week build timeline
Four Conversation Types
  • Job Interview Practice

    Answering questions about experience, strengths, teamwork, and problem-solving

  • Workplace Communication

    Handling feedback, conflict resolution, supervisor communication

  • Parole Hearing Preparation

    Practicing common questions and effective response frameworks

  • Reentry Planning

    Identifying needs for housing, ID, transportation, healthcare, employment

Key Constraints

Prison Environment
  • • 15-minute maximum call duration with 30-minute cooldown between calls
  • • Must work reliably over prison phone systems (GTL/Securus/Viapath)
  • • Risk of number blacklisting if content flagged
Security & Privacy
  • • Cannot maintain memory between calls (security concern about coded communication)
  • • Flexible data storage based on jurisdiction (some facilities: metadata only; others: full transcripts)
  • • PII redaction capabilities required
Technical Requirements
  • • Automatically accept collect calls and confirm charges
  • • CEO absorbs all call costs (no financial burden on users)
  • • Strict security & compliance (FCC, CJIS standards)

Success Criteria

Technical Metrics

99.5%

Uptime

<3s

One-way latency

90%

Connection rate

<5%

Dropped calls

Operational Success
  • Works reliably under real prison conditions
  • Coaches can access usage data and insights
  • Secure, compliant, and ethical by design

Section 3

Technical Approach

Proposed Architecture

Prison Phone System

Twilio Voice API

AWS Processing Layer

Voice AI Engine

Response to Caller

Core Technology Stack

Telephony: Twilio
  • • Collect call handling with automated DTMF acceptance
  • • Dynamic number allocation per facility
  • • Programmable voice for disclaimers
Speech Processing: ElevenLabs
  • • Integrated STT/TTS optimized for telephony
  • • Handles poor audio conditions gracefully
  • • English and Spanish support
LLM Engine: Claude
  • • Conversational AI with safety features
  • • Nuanced coaching capability
  • • Content moderation guardrails
Infrastructure: AWS
  • • End-to-end encryption (TLS 1.3, AES-256)
  • • GovCloud available for compliance
  • • Scales from 25 to 200+ concurrent calls

Key Technical Solutions

1. Phone as User Interface
  • • Users verbally select coaching modules: "I want to practice job interviews"
  • • Each session is a standalone 15-minute chunk
  • • No memory between calls (security requirement)
  • • Graceful wind-down at 13-14 minutes to prevent hangups
2. Collect Call Handling
  • • System automatically detects carrier prompt: "You have a collect call from [Name] at [Facility]"
  • • Sends DTMF "1" to accept charges
  • • Prepaid account covers all costs
  • • Optional: Extract facility name, discard personal identifiers
3. Flexible Data Storage
  • Level 1: Metadata only (timestamp, duration, facility, technical metrics)
  • Level 2: Metadata + anonymized transcripts with PII redaction
  • Level 3: Metadata + transcripts + facility-approved identifiers
  • • Configurable per jurisdiction requirements
4. Leveraging Your Existing System
  • • Leverage your proven VAPI prompts
  • • We add: collect call automation, prison-grade reliability, enhanced security
  • • 25+ concurrent calls (vs. 5 in pilot)

Curriculum Retrieval Flow

The system intelligently guides users to the right coaching module through natural conversation, adapting to whether they know exactly what they want or need help finding their place.

Curriculum retrieval flowchart showing user paths from call connect through module selection to coaching session

Visual Flow Example

Curriculum retrieval flow showing a user calling CoachLink, selecting a module, and beginning a coaching session

Full call flow from connection (0:00) to coaching session start (0:30)

Addressing Prison Constraints

ChallengeOur Solution
Multi-second latencyStreaming responses, intelligent buffer management
Poor audio qualityNoise cancellation, high confidence thresholds, graceful fallbacks
30-minute cooldown penaltyUX designed to prevent hangups—clear exit points, encouraging language
Number blacklisting riskStrong content guardrails, facility-specific number pools, monitoring
No cross-session memoryStateless sessions by design, verbal navigation, module-based curriculum

Section 4

Timeline & Investment

Build Timeline: 8 Weeks

Week 1

Discovery & Planning

  • • Migrate existing coaching prompts
  • • Finalize data storage policies per facility
  • • Design 15-minute session flows
Weeks 2-5

Development

  • • Twilio infrastructure + collect call automation
  • • Voice AI pipeline (ElevenLabs + Claude)
  • • Dynamic number allocation
  • • PII redaction system
Weeks 6-7

Testing & Refinement

  • • Hands-on testing by CEO team
  • • Collect call simulation
  • • Latency optimization
  • • Iterative improvements
Week 8

Launch Prep

  • • Final polish
  • • Documentation
  • • Facility deployment guides
  • • Knowledge transfer

Investment

BUILD

$19,000

PROJECT-BASED

  • • 8-week intensive development
  • • All PRD deliverables
  • • 25 concurrent call capacity
  • • Collect call acceptance system
  • • Flexible data storage + PII redaction
  • • Complete documentation and training

Section 5

Next Steps

1
Vendor Selection

You're evaluating vendors and will select 2-3 for follow-up conversations.

2
Follow-Up Call

More detailed technical discussion and address remaining questions.

3
Kickoff

Target: January 6-10, 2025. Begin Week 1 discovery with access to existing prompts and materials.

We're ready to start in early January and deliver a production-ready system in 8 weeks.

Thank you for considering us as your technical partner.

Let's build something great together.