Hokma Solutions / Case study

Approval queue and agent router

Personal operations platform. Live, internal use only, not a commercial product.

My role. System architect and operator

Stack: Next.js, React, Supabase Realtime, Python 3, Claude scheduled tasks, Gemini Flash, xAI Grok, GitHub Issues API

Decide the cards yourself

Agents propose. You approve, edit or reject. Each decision is checked against the agent's sealed guess and written to the journal.

Animated sample. Every name, task and number in it is invented.

Approvals4 waiting

Outreach agentEffect: Medium

Send a follow-up to 12 trial sign-ups

Hi Sam, you started a trial last week. Most teams get value from the weekly report first. Want me to set it up for you?

Source
Trial list exported Tue 09:10
Model
Small model

The agent's guess is sealed until you decide

An agent cannot do these alone

  • Send an emailCard
  • Spend moneyCard
  • Publish a postCard
  • File a documentCard

Decision journal

Decisions
0
Edited
0%
Agent guessed right
0 of 0
  1. Decide a card and it lands here.

The pace router

Each AI subscription has a weekly limit that resets at a different time. The router sends work where spare room is worth the most: lots left, and a reset coming soon. The thin line on each bar is where usage would be on an even pace.

  • Claude81% used · resets in 30h
  • ChatGPT44% used · resets in 110h
  • Gemini22% used · resets in 140h
  • Grok58% used · resets in 9h

Send a task to see where it goes and why.

The problem

Running specialized workflows across several AI providers (Claude, OpenAI, Gemini, Grok) usually leads to scattered context, quotas running out mid-week, prompt-injection exposure from untrusted web pages and emails, or unattended actions nobody approved.

How it was built

  • Designed a schema for workspaces, members, providers, connections, agents, runs, usage snapshots, tasks and lessons, with row-level security.
  • Built the two-tap approval card protocol: no agent can send an email, spend money, publish a post, or file a document. The agent writes a card with the draft, the source, the model tier, and a decision prediction (approve, edit or reject) that stays hidden until the person decides. Every decision goes into a journal so edit rates can be measured.
  • Built a multi-provider pace router that tracks weekly subscription usage across Claude, ChatGPT, Gemini and Grok against their reset times, so work goes to accounts with spare headroom.
  • Built a cross-assistant task board on security-definer database functions, so assistants can hand off research and browser tasks without gaining approval rights.
  • Tested a multi-agent outreach loop between two assistants using verified GitHub issue comments, with scripts that keep personal and business data apart.

Results I can show

  • Scheduled agent roles run under version-controlled prompts with evaluation suites.
  • No routine can send, spend or publish without an approval card.

Limits

  • Built for a single owner first. It is shown here as proof of how I design approval and routing, not as a product for sale.
  • Subscription usage meters need a scheduled browser reader, because Anthropic, OpenAI, Google and xAI do not offer APIs for consumer subscription usage.

Try the demos

Synthetic data. The local rules engine is the real deterministic code. The AI layer here returns sample output. No model is called.

A simulation of the two-tap approval card and the multi-provider pace router, on synthetic tasks. The agent's prediction stays hidden until you decide.

Pace router

  • Anthropic (Claude Sonnet): 54% used this week, resets in 52h (behind pace (-14 pts))
  • Google (Gemini Flash / Gem): 38% used this week, resets in 52h (behind pace (-30 pts))
  • OpenAI (gpt-4o-mini API): 79% used this week, resets in 120h (ahead of pace (+11 pts))

sc-sender · effect: external (medium)

Send today's outbound batch: 4 verified prospects

The agent's prediction is committed and hidden until you decide.

Fifteen minutes, and you will know if it's a fit.

Tell me what your team is dealing with. I will tell you whether I can help and what it would cost.

Book a 15-minute call

Prefer to write?

Send a short brief