Johny Patrick Hokma Solutions · AI & Support Systems
← All Consulting Work / Project Deep Link for Applications
Internal Agentic Platform (Live) Personal Operations Platform · Live Internal Dogfooding (Not a Primary Commercial Product)

Life OS & Acquisition Loop — Human-in-the-Loop Multi-Agent Operations Platform

My Role: System Architect & Operator

Next.js 16React 19Supabase Realtime & VaultPython 3 (bridge/sync.py)Claude Scheduled TasksGemini FlashxAI GrokGitHub Issues API
Command Center PWA (Private Auth Gate) ↗ Scope a Similar Project →
Interactive Capability Proof · Synthetic Data

Test the Actual Engines Behind the Case Studies

Select any workbench tab below, click a preset scenario or enter your own text, and inspect the live deterministic output. All demonstrations use synthetic data and clearly distinguish live local algorithms from simulated LLM responses.

Mode Disclosure: Layer 1 Local Rules Engine executes LIVE deterministic code (0ms, zero tokens). Layer 3 Pre-Send 8-Dimension Check runs in Labeled Sample / Simulated AI Mode because OPENAI_API_KEY and PUBLIC_CHECK_ENABLED=true are kept off in this public sandbox.
Direct Link: https://consulting.supportcoach.io/projects/support-coach-extension →
Privacy-Verified Outbound Payload (apps/extension/src/lib/apiClient.ts)

Customer messages never leave the browser. Only these 4 fields cross the wire:

{
  "draft_text": "Like I said, you need to reconnect it yourself. It's not a bug on our end, and there is nothing I can do from here.",
  "triggered_rules": [
    "dismissive_harsh",
    "blame_language",
    "missing_ownership",
    "cold_transactional"
  ],
  "local_score": 72,
  "draft_hash": "ead21245a262ef2d"
}
LAYER 1 · LOCAL RULES ENGINE (0ms · NO AI) Local Risk Score: 72/100
This could read as dismissive or abrupt to the customer. INSTANT HINT · NO AI
Example: "I can help get this working for you right now — let me walk you through the steps."
Accept behavior: Dismiss only (presence families require AI rewrite to preserve specific facts)
Triggered Rule Families: dismissive_harsh, blame_language, missing_ownership, cold_transactional
LAYER 3 · PRE-SEND CHECK (SAMPLE / SIMULATED AI MODE) HIGH RISK

What's working: Identifies the reconnection requirement needed to resolve the issue.

empathy2/10
ownership2/10
clarity9/10
actionability9/10
professionalism2/10
de escalation2/10
conciseness9/10
Suggested Rewrite (WYSIWYG · Fact-Preserving)

I can help you get this reconnected. This happens when the local connection token resets rather than a platform outage, and I can walk you through reconnecting it right now so everything syncs cleanly.

Mode Disclosure: Demonstrates the exact deterministic parser, remote screen-share gap exemption (≥5 min with session URL), abandoned-chat neutral scoring (7/10), and Product Confusion Topic schema from supportcoach-ai (Tasks 8, 33, 37–40) on synthetic transcripts.
Direct Link: https://consulting.supportcoach.io/projects/support-coach-dashboard →
Tip: Test Scenario B to see how a 16-minute screen-share gap is exempted from response-time penalties.
QA ENGINE & EDGE-CASE GUARDRAILS Attention: HIGH
  • Orphan Line & Sender Normalization (Task 8e/8h): Multi-line messages preserved under active speaker.
  • Response-Time Threshold Rule: Gaps under 2 min ignored; only gaps >4 min without screen-share URLs are cited (max 2-3 timestamps).
Manager Copy-Paste Coaching Feedback

Quick note on this chat: you got the customer to the right place (Settings > Templates), which solved the markup question. For next time, lead directly with what the product CAN do ("You can hide markup by switching the template in Settings > Templates") rather than opening with "That is not possible", and skip phrases like "As I said earlier" after a 4.1-minute gap.

Empathy & Workaround Ownership Second occurrence in 30-day window (Manager override takes precedence)

Observed (stays in dashboard): Agent replied 'That is not possible on the standard invoice view' and 'As I said earlier' before guiding the customer to the Templates setting.

Recommended (eligible for bridge): Lead with the available path ('You can hide line-item markup using a Client Template under Settings > Templates') instead of opening with a flat 'not possible', and avoid 'As I said earlier'.

Response Pacing (>4 Min Threshold) New coaching point

Observed (stays in dashboard): Wait of 4.1 minutes (10:03:48 to 10:07:55) without an interim check-in.

Recommended (eligible for bridge): When researching a tax or template question takes longer than 3 minutes, send a brief hold note so the customer knows you are still checking.

PRODUCT CONFUSION TOPICS EXTRACTED (TASKS 37–39)
Invoices › email invoice without showing line-item markup
Category: confusion · Confidence: 94% · Status: matched_approved_topic
"Hi, I'm trying to email an invoice to my client without showing the line-item markup column, and the PDF template keeps putting the old tax total back."
Invoices › email invoice without showing line-item markup
Category: confusion · Confidence: 94% · Status: matched_approved_topic
"Wait, isn't there a client-facing template option where markup is hidden? And why did the tax revert?"
Mode Disclosure: Simulates the multi-provider pace router (002_platform.sql) and two-tap Approval Card protocol (CLAUDE.md & journal/README.md) on synthetic agent tasks. Notice how the agent's decision prediction stays hidden until you tap Approve, Edit, or Reject.
Direct Link: https://consulting.supportcoach.io/projects/agentic-governance-ops →
MULTI-PROVIDER PACE ROUTER
Anthropic (Claude Sonnet) behind pace (-14 pts)
54% used this week · resets in 52h · capabilities: browser_pc scheduled_runs sql_connector
Google (Gemini Flash / Gem) behind pace (-30 pts)
38% used this week · resets in 52h · capabilities: web_research voice_schedule
OpenAI (gpt-4o-mini API) ahead of pace (+11 pts)
79% used this week · resets in 120h · capabilities: structured_json low_latency_check
NEEDS YOUR YES · TWO-TAP APPROVAL CARD sc-sender · effect: external (medium)

Send today's outbound batch: 4 verified prospects

Agent prediction is committed and hidden until you tap Approve, Save Edit, or Skip above (per Constitution Rule: "Write the prediction before he acts and keep it hidden until he has").

1. The Client & Operational Problem

Running specialized workflows across multiple AI providers (Claude, OpenAI, Gemini, Grok) typically leads to context fragmentation, uneven quota exhaustion mid-week, prompt-injection exposure from untrusted web pages/emails, or dangerous unattended actions.

2. Architecture & Implementation Evidence

  • Architected a multi-tenant-ready schema (db/001_docs.sql to 004_tasks_board.sql: workspaces, members, providers, connections, agents, runs, usage_snapshots, assistant_links, tasks, lessons) with Row-Level Security.
  • Built the Two-Tap Approval Card protocol (queue/<id>): no agent can send an email, spend money, publish a post, or file a document. Instead, the agent writes a card containing the draft, source, model tier, and a pre-committed decision prediction (approve / edit / reject) that stays hidden in the UI until the human acts. Every decision is logged to a structured decision journal (journal/README.md) to measure edit rates.
  • Built a Multi-Provider Pace Router & Accounts dashboard (accounts/page.tsx) that tracks weekly subscription usage across Claude, ChatGPT, Gemini, and Grok against their reset timestamps so work is routed to accounts with spare headroom.
  • Created a cross-assistant Task Board (/hand/[token]) using security-definer Postgres RPCs (board_me, board_list, board_update, board_ask, board_lesson) so Grok, Gemini, and Claude can hand off research and browser tasks without gaining approval privileges.
  • Verified a multi-agent B2B acquisition loop (support-coach-acquisition) between Grok and Claude using nonce-verified GitHub issue comments (ACQ-TEST-001 passed Nov 7, 2026) and strict data-wall scripts (bridge/sync.py).

3. Demonstrated, Verifiable Results

  • 10+ scheduled cloud and PC-bound agent roles running under version-controlled prompts (roles/*.md) with explicit golden-day evaluation suites (evals/life-chief.md).
  • Strict cryptographic and script-level separation between personal, business (Hokma Solutions), and employer contexts.
  • Zero unattended external sends or unauthorized spend across all active routines.

4. Honest Boundaries & Limitations

  • Built as a single-owner command center first (Life OS is showcased here strictly as architectural proof of multi-agent governance and Next.js 16 + Supabase Realtime engineering, not as the primary commercial product).
  • Provider subscription usage meters require a scheduled browser reader task because Anthropic, OpenAI, Google, and xAI do not offer REST APIs for consumer/pro subscription usage.
Evidence-Backed Consulting Offers

Three Concrete Service Offers Mapped to Buyer Needs

Grounded in working codebases and verified against current fixed-price market demand — without forcing distinct capabilities into a single narrow niche.

Offer 01 · Support Operations & CX AI

Customer Support AI, Pre-Send Guardrails & Transcript QA Systems

Who it helps:

SaaS founders, Heads of Support, and CX Operations Leaders running 5–50 agent support teams on Zendesk, Intercom, Zoho SalesIQ, or custom help desks.

Problem it solves:

Standard AI chatbots hallucinate policies or frustrate customers, while post-hoc QA tools ($35+/agent/mo) only tell managers on Thursday about a live chat that escalated on Monday. Meanwhile, generic transcript analyzers penalize agents unfairly for screen-share gaps or abandoned chats.

Concrete Deliverable:

A tailored Support QA & Coaching Pipeline or Pre-Send Reply Checker: (1) Custom rubric & company tone-profile calibration, (2) Domain-aware transcript parser handling multi-line/screen-share/abandoned chat edge cases, (3) Product Confusion Topic clustering with verbatim customer quotes, (4) Manager follow-through scorecards & Slack/email PDF coaching handouts, and (5) Hard per-seat OpenAI spend caps.

Supporting Repository Credibility:
Requires Further Validation Per Client:

If direct API ingestion is required for help desks other than Zoho SalesIQ (e.g., Zendesk Tickets API, Intercom Conversations API, or Shopify/Re-amaze), a 2–3 day API payload & rate-limit discovery spike is required before locking fixed-price ingestion scope.

Scope This Offer →
Offer 02 · In-Page Browser Copilots

Manifest V3 Chrome Extensions & Hybrid Local + AI Overlay Copilots

Who it helps:

B2B SaaS products and operations teams that need real-time AI assistance, compliance checks, or workflow automation injected cleanly inside third-party web applications (Zendesk, Intercom, Zoho, Salesforce, HubSpot).

Problem it solves:

Third-party web apps use complex, shifting DOMs (contentEditable vs textarea, dynamic modals, Shadow DOMs) where naive browser extensions lose state on chat switch, get hidden behind overlays, break native Send buttons, or burn LLM API budget on every keystroke.

Concrete Deliverable:

A production-ready Manifest V3 Chrome Extension (TypeScript + Vite) + Node.js/Express Backend featuring: (1) Modular platform-adapter DOM layer with synthetic event dispatch and edit-box focus guards, (2) Instant 0ms local rules engine for deterministic first-pass detection, (3) Debounced/on-demand AI rewrite cards in isolated Shadow DOM, (4) Passive UI self-diagnostics, and (5) Chrome Web Store package & IT security disclosure page.

Supporting Repository Credibility:
Requires Further Validation Per Client:

Target host web application DOM inspection (confirming whether composers use contentEditable, iframes, or closed Shadow DOMs, and verifying whether operators use the browser or a native desktop app).

Scope This Offer →
Offer 03 · Agentic Workflows & SaaS MVPs

Human-in-the-Loop Multi-Agent Workflows & Full-Stack Next.js + Supabase MVPs

Who it helps:

Founders, agencies, and technical operators who want Claude/OpenAI/Gemini automation or a full-stack SaaS MVP where every external action (sending emails, updating records, spending budget) is gated by a fast human approval card.

Problem it solves:

Unattended AI agents hallucinate, leak private context across accounts, or send off-brand messages, while purely manual workflows bottleneck the founder. Teams also struggle with unauthenticated LLM routes and missing Row-Level Security (RLS) in rapid vibe-coded prototypes.

Concrete Deliverable:

A governed AI Workflow or Full-Stack SaaS MVP (Next.js App Router + Supabase PostgreSQL with RLS + Claude/OpenAI): (1) Versioned role prompts & evaluation test suites, (2) Mobile-ready two-tap Approval Queue (Approve / Edit / Reject) with decision journaling, (3) Multi-provider router & spend guardrails, or (4) Production security & RLS hardening audit of an existing Next.js/Supabase codebase.

Supporting Repository Credibility:
Requires Further Validation Per Client:

Review of client's target LLM provider terms (API keys vs native scheduled tasks) and existing database schema/RLS posture before deployment.

Scope This Offer →
Project Enquiry & Fixed-Price Scope Route

Scope a Fixed-Price Milestone or Technical Audit

Use this interactive scope builder to submit a project enquiry, or book a 15-minute discovery call directly on Calendly.

Enquiry Route Configuration Status:
  • Local Structured Persistence (ACTIVE): Submitting this form validates your scope and saves a timestamped record to data/enquiries.json on this server, generates a downloadable Markdown/JSON project brief, and builds a pre-filled mailto: link.
  • External Transactional Email (UNCONFIGURED IN SANDBOX): Direct automated email delivery via Resend REST API requires setting RESEND_API_KEY and NOTIFY_FROM_EMAIL in environment variables. We explicitly label this rather than faking an outbound email send.
  • Live Scheduling Link (AUTHORIZED): calendly.com/supportcoach-support ↗

Direct Channels & Verification

Calendly Discovery Call:
https://calendly.com/supportcoach-support ↗

Live Commercial Products:

Saved Local Enquiries (0)

Submit a scope brief on the left to inspect the structured payload and download a ready-to-attach Markdown brief for Upwork or email.