J
JEREMY
"That friend who reads the fine print."

Sovereign AI legal agent for the Pro Se Network. 80 phases built. 62 API endpoints. 22 tools. 4-guard governance. Trained on 5,196 legal Q&A pairs across 12 areas of law. Total build cost: $0.00. LIVE at prosenetwork.org — Alpha.

PRO SE NETWORK
Visit Site Explore the Engine
01 / Meet Jeremy

The Numbers

Built in 80 phases. Zero shortcuts. Zero dollars.

80
Phases Built
62
API Endpoints
5,196
Training Pairs
$0
Build Cost
"I exist because two men decided that the law should belong to everyone — not just those who can afford it."
02 / The Numbers

The Problem

100 million Americans cannot afford legal representation. The system doesn't pause for them.

📄

Form Fillers

LegalZoom, Rocket Lawyer. Fill in blanks. No procedural guidance. No risk assessment. No deadlines.

🤖

Generic AI

ChatGPT hallucinates case law. No jurisdiction awareness. No guardrails. No consequences when it's wrong.

Legal Aid

Underfunded. Months-long waitlists. Limited areas of law. Not enough attorneys for 100M people.

The gap isn't information — it's navigation. People don't need to read the law. They need someone to walk them through it.

03 / The Problem

The Engine

Five layers. Deterministic first, AI second. The LLM cannot override the rules.

Layer 5: Attorney Referral Network — The Off-Ramp Layer 4: Deliverables Engine — Documents, Walkthroughs, Timelines Layer 3: Risk & Safety Engine — Scoring, Gates, Circuit Breakers Layer 2: Legal Intelligence Engine — Rule Database (The Moat) Layer 1: User Intake & Case Assessment — Deterministic State Machine

State Machine

Every case flows through defined states with deterministic transitions only. No LLM decides transitions. Skip a state? InvalidTransition.

INITIAL → NARRATIVE_INTAKE → FACT_EXTRACTION → JURISDICTION_DETECTION → AREA_CLASSIFICATION → RISK_ASSESSMENT → PREREQUISITE_CHECK → DEADLINE_CALCULATION → GUIDANCE_READY → DOCUMENT_DRAFTING → USER_REVIEW → COMPLETE

Rule Database

This is the moat. Not templates — rules. Prerequisites with blocking flags. Deadlines with strict enforcement. Causes of action with elements.

Title VII Prerequisites: 1. EEOC Charge Filed (BLOCKING) — 180/300 days from incident 2. Right-to-Sue Letter (BLOCKING) — Request after 180 days 3. File within 90 days of letter (STRICT) System WILL NOT generate complaint until both prerequisites confirmed.
04 / The Engine

The AI Brain

Three-tier fallback. Free first. Always.

TierModelCostTokensUse Case
Tier 1Gemini 2.5 Flash Lite$0.001,000Simple chat, quick answers
Tier 2Gemini 2.5 Flash$0.001,500Tool sessions, medium complexity
Tier 3Gemini 3.1 Pro Preview~$0.012,000Complex legal structured tasks

Auto Tier-Up

  • Complex tasks → Tier 2 minimum
  • Input > 2,000 chars → Tier 2
  • Tool follow-ups → Tier 2
  • Failed response → Step up
  • All exhausted → Graceful fallback

Fine-Tuned Model

Jeremy also has a purpose-built Phi-3 Mini 3.8B on HuggingFace. 32 attention heads. 3,072 hidden dimensions. Trained on 5,196 legal Q&A pairs. Proof that fully local, zero-cost legal AI is possible.

PHI-3 3.8B 7.64 GB $0 TO TRAIN
05 / The AI Brain

Risk Scoring

Five signals. One composite score. When it hits CRITICAL, Jeremy stops himself.

0 – 25
LOW — Full Guidance
25 – 50
MEDIUM — With Warnings
50 – 80
HIGH — Limited + Urged
80 – 100
CRITICAL — STOPPED
SignalImpact
Custody DetectionInstant CRITICAL (score: 100)
Area-of-Law Default5 – 50 points
Deadline Urgency Triggers+10 per trigger
Prerequisite Gaps+5 per missing
Aggravating Factors+10 – 20 per factor (minor, DV, deportation, foreclosure)
"When risk hits CRITICAL, all AI guidance stops. Jeremy routes exclusively to attorney referral. This is a constitutional constraint, not a feature."
06 / Risk Scoring

The Toolbox

Jeremy doesn't just talk — he acts. 22 tools. 4 risk levels.

Low Risk (Auto)

  • Read files & directories
  • Web search
  • Read SMS inbox & threads
  • Search email
  • Memory search & read
  • Task list & comment
  • GitHub pull

High Risk (Guard Vote)

  • Execute shell commands
  • Send email (draft → owner → send)
  • Send SMS (draft → owner → send)
  • GitHub push
  • Resolve tasks

Critical (Full Gate)

  • Request changes to own capabilities
  • Propose to governance guards
07 / The Toolbox

The Governance

Jeremy doesn't govern himself. Four guards. One vote.

💰

West

Financial & Risk

Gemini
🛡

East

Physical & Safety

Kimi
📚

North

Advisory

GPT
💻

South

Technical

Claude
Jeremy requests HIGH tool ↓ Intent classifier routes to guard(s) ↓ All 4 guards VOTE (APPROVE / DENY / ABSTAIN) ↓ Majority required to proceed ↓ Owner holds SUPREME OVERRIDE
"Nothing sends without owner review. Text, email, any communication. Draft → present → owner approves → then send. This is law."
08 / The Governance

The Voice

Jeremy speaks through Carter — warm, charismatic, New York edge. His voice is alive — a 32-bar KITT equalizer pulses gold with every word.

Personality

  • Sharp, charismatic, warm
  • Autonomous — "not under guidance"
  • Direct — no stiff formality
  • Honest about limits — never hallucinate
  • Developer-aware — reads code + law

Voice Specs

PropertyValue
ProviderElevenLabs
VoiceCarter
Modeleleven_flash_v2_5
PlanStarter ($5/mo)
Quota30,000 chars/mo
Delivery/api/tts (500 char cap)

KITT Equalizer

32 gold bars react to Jeremy's actual voice via Web Audio API FFT analysis. Center bars glow brighter gold, edge bars fade to brass. Active bars emit gold box-shadow glow. Idle state breathes with a slow pulse.

32 BARS FFT-128 REAL-TIME

Clean TTS Pipeline

Markdown is rendered as styled HTML in chat (headings, bold, lists). Before voice synthesis, cleanForTTS() strips all markdown, HTML tags, and formatting — so Carter speaks clean English, never "hashtag hashtag hashtag."

"Jeremy is not designed to sound like a lawyer. He's designed to sound like the smartest person in your corner who happens to know the law."
09 / The Voice

Proof of Work

Jeremy is not theoretical. He is supporting real cases in real courts.

McDaniel v. City of New York

Section 1983 — False Arrest + Excessive Force by NYPD

ElementDetail
JurisdictionNY Supreme Court, Bronx County
Complaint90 paragraphs, 4 causes of action
Valuation$450,000 – $650,000
Evidence32 exhibits catalogued
Key FindingOfficers changed charge from Theft of Services to Resisting Arrest to cover assault
ProofTransit card shows 2 fare payments — irrefutable electronic evidence
MedicalBellevue CT confirmed concussion
REAL CLIENT REAL COMPLAINT REAL COURT REAL STAKES
10 / Proof of Work

The Cost

$66.88 per year. That's it. That's the whole thing.

ComponentChoiceCost
Development (80 phases)Blood, sweat, code$0.00
Training (GPU)Kaggle free quota$0.00
Primary VMOracle Micro 2 (free tier)$0.00
Backup VMGCP (free tier)$0.00
AI ModelsGemini free tier + Phi-3 open-source$0.00
Legal DataCourtListener (open source)$0.00
SSL CertificateLet's Encrypt (auto-renew)$0.00
Domainprosenetwork.org (Porkbun)$6.88/yr
VoiceElevenLabs Starter$60/yr
TOTAL ANNUAL COST$66.88
$0
Build Cost
$0.01
Per Case (AI)
$5.57
Per Month Total
11 / The Cost

What's Different

Not another chatbot with a legal prompt.

FeatureGeneric Legal AIJeremy
Core EngineLLM with legal promptDeterministic state machine + LLM interface
Hallucination Control"Please don't hallucinate"LLM cannot override rule engine
Risk ManagementNone5-signal composite scoring + circuit breakers
Cost$20-100+/mo$0 to build, ~$0.01 per case
GovernanceNone4-guard Quadrant Guard consensus voting
MemoryConversation context onlyPersistent FTS5-indexed knowledge base
VoiceText onlyElevenLabs Carter + KITT equalizer
Data FlywheelNo learningConversation logging (JSONL) for retraining
DeploymentDemo / localhostLIVE at prosenetwork.org (SSL, Oracle VM)
Real CasesDemos onlyActive cases, real clients, real courts
12 / What's Different

The Roadmap

Phase 1 is done. Here's where we're going.

Phase 1 — Complete

80-phase legal engine. 62 endpoints. 22 tools. Quadrant Guard. Fine-tuned model. Voice. Memory. Active case support. Attorney referral network. Divorce module.

Phase 2 — In Progress (Alpha Live)

Completed: Live deployment (prosenetwork.org + Let's Encrypt SSL), unified dark-gold UI across all pages, KITT 32-bar audio equalizer, conversation data logging (JSONL) for retraining, markdown rendering, clean TTS pipeline, divorce guided workflow. Next: Additional jurisdictions, inference optimization, 10,000+ training pairs, mobile app.

Phase 3 — Growth

50-state coverage. Multi-LLM hallucination detection. Real-time court rule updates. Immigration deep module. Criminal defense deep module. Third-party API.

Phase 4 — Scale

Multilingual expansion. Attorney marketplace. White-label for legal aid organizations. Jeremy as a platform.

12 Areas of Law12 / 50 states
Training Data5,196 / 10,000 pairs
Phase 1 Completion80 / 80 phases
13 / The Roadmap
THE MISSION
"Because everyone deserves to understand the law that governs them."

Jeremy is not just software. He is an access-to-justice infrastructure capable of reshaping how ordinary people navigate the law. The deterministic rule engine ensures reliability. The AI layer adds understanding. The attorney network provides human oversight. The $0 infrastructure model ensures sustainability.

Visit Pro Se Network Technical Whitepaper Back to Top
PRO SE NETWORK PROSENETWORK.ORG ALPHA LIVE

Built by Israel "Ace" Burns & Esco Obong

Esuk Atu — Calabar — New York

14 / The Mission