SoulShield App

A trauma informed AI support for survivors of human trafficking and domestic violence, offering bystander-safe, neutral support during nights and between therapy sessions when help is unavailable

Comment

Inspiration

It started with a conversation that changed everything. During a meeting with a survivor support organization in Austin, a clinical staff member told us:

"If it happens after hours or on the weekend, we're obviously not here for that. The survivors do have a lot of emotional distress, especially in the first few weeks."

We learned that 50 million people are trapped in modern slavery worldwide. Those who escape receive therapy once a week — but trauma doesn't follow a schedule. Panic attacks happen at 3 AM. Nightmares don't wait for business hours.

Existing mental health apps fail survivors for one critical reason: they use obvious clinical language. If an abuser checks a victim's phone and sees words like "therapy" or "abuse," it could mean violence. We asked: What if we could build support that hides in plain sight?

What it does

SoulShield is a trauma-informed AI support companion that helps survivors navigate emotions between therapy sessions. It operates in three modes:

  1. Grounding Mode — Box breathing, 5-4-3-2-1 sensory exercises when overwhelmed
  2. Emotion Exploration — Helps name feelings without giving advice
  3. Session Preparation — Organizes thoughts for the next therapy appointment

Key innovation: Bystander-Safe Mode — When users indicate someone might be monitoring them ("He checks my phone"), SoulShield automatically uses neutral wellness language that won't raise suspicion. It looks like a generic mindfulness app. SoulShield is not just a bystander tool — it’s designed for survivors’ everyday emotional regulation between therapy sessions, when no one else is available.

Additional features:

  • Voice input/output for users who can't safely type
  • Dark mode for nighttime use
  • 24/7 availability
  • Crisis resources (hotlines)
  • Enterprise-grade security with client-side encryption

How we built it

Multi-Agent AI Pipeline:

Agent Purpose
Routing Agent Classifies message by mode, privacy context, and risk level
Support Coach Generates trauma-informed responses using user's words
Safety Fallback Catches errors with calming, generic responses

Tech Stack:

  • Frontend: Vanilla JS on Vercel with Web Speech API for voice
  • Backend: AWS Lambda + API Gateway + DynamoDB
  • AI: OpenAI GPT-4 with prompts managed in LangSmith
  • Security: 5 layers including AES-256 client-side encryption

Prompt Engineering: We iterated dozens of times. The breakthrough came when we stopped asking AI to "be a therapist" and instead asked it to "be a calm presence that helps someone breathe."

Challenges we ran into

  1. "Robotic Therapist" Problem — Early responses felt scripted and unnatural.

    • Solution: Removed prescriptive rules, added instructions to mirror the user's words naturally.

  2. Bystander Detection — Users don't say "someone is watching." They say "he checks my phone."

    • Solution: Created evaluation dataset with 40+ edge cases to train subtle signal detection.

  3. Security for Vulnerable Users — Standard security isn't enough when a breach could be life-threatening.

    • Solution: Implemented 5 layers including zero-knowledge encryption (we can't read messages).

  4. Mobile Experience — Survivors use phones in difficult conditions.

    • Solution: Complete mobile-responsive redesign with voice input for when typing isn't safe.

  5. Validation Without Users — We're technologists, not trauma experts.

    • Solution: Multiple stakeholder meetings with clinical staff who confirmed this addresses a real need.

Accomplishments that we're proud of

  • Bystander-Safe by Design: Zero instances of dangerous language when privacy is compromised
  • Stakeholder Validation: A survivor support organization confirmed this addresses a real, unmet need
  • Production Deployment: Live at soulshield.vercel.app — not just a prototype
  • Enterprise Security: 5-layer protection including client-side encryption and auto-delete
  • Multi-Agent Architecture: Three specialized AI agents working in sequence for reliability
  • Voice Support: Full speech-to-text and text-to-speech for users who can't safely type
  • Prompt Engineering Breakthrough: Responses that feel human, not robotic

What we learned

From Trauma Experts:

  • Grounding techniques are the first line of defense against panic
  • AI should supplement, never replace the therapeutic relationship
  • Sometimes just having a calm presence at 3 AM is enough

From Building:

  • Multi-agent architectures allow specialization and fault tolerance
  • Prompt engineering is an art — subtle wording changes transform "robotic" into "therapeutic"
  • LangSmith transforms development with version-controlled prompts and evaluation datasets

From Testing:

  • Bystander-safe language passes the "over-the-shoulder" test
  • Voice input is crucial for some users
  • Dark mode isn't aesthetic — it's safety at night

What's next for SoulShield App

Q1 2026:

  • Guided grounding exercises with visual timers
  • Offline mode (works without internet)
  • Feelings wheel integration

Q2 2026:

  • Therapist dashboard (with user consent)
  • Pattern recognition for therapy preparation
  • Multi-language support

Future:

  • Clinical validation research partnerships
  • HIPAA compliance certification
  • White-label version for organizations

*We didn’t just build an app — we built a protective bridge for the 167 hours between therapy sessions.

Built With

  • aes-256-encryption
  • aws-api-gateway
  • aws-cdk
  • aws-dynamodb
  • aws-lambda
  • cursor
  • html/css
  • javascript
  • langchain
  • langsmith
  • openai-gpt-4
  • python
  • vercel
  • web-speech-api
Share this project:

Updates

Submission history