JD
All projects

AI Employees Platform

Heynet

A team of AI employees — for your phone and for your company.

Founder — product, prompt architecture, and growthLive product · iOS + Web + SMS

01

Problem

Two versions of the same problem, at opposite ends of the day. Personally, your phone works against you: spam calls interrupt, voicemails go unheard, and the small tasks — a reminder, a quick answer, is this thrift-store find worth $40 — never justify opening an app. Professionally, every AI tool starts from zero, so you re-explain your company, your customers, and your voice to a blank chat box every single time.

Both are staffing problems dressed up as software problems. The answer isn't another chatbot — it's employees: specialized roles with shared memory, real channels, and enough context that you stop briefing them.

02

Solution

Heynet ships as two products on one substrate. Heynet Personal handles the phone and the day: an AI Call Screener that answers unknown callers, an SMS assistant you can text without opening anything, an AI Reseller that appraises items from a photo, and proactive daily briefings. Heynet Work is company intelligence: a directory of 55,000 enriched company profiles, private workspaces per company, and AI workers that draft with your real context.

The connective tissue is the Knowledge Trainer. Upload documents, notes, and corrections once, and every employee — the screener writing a call summary, the email assistant drafting a follow-up — answers with your details in your tone. Train once; the whole staff gets smarter.

Everything writes back to one memory layer. Call summaries land in chat history and vector memory, so an employee that had nothing to do with the call can still reference it next week. That's the difference between a set of AI tools and an AI team.

Capabilities

What it does

AI Call Screener

Answers unknown callers with a greeting written per call from your name, the time of day, and your live calendar — then records, transcribes, summarizes, and texts you who called and why.

AI Messenger

Text your AI from anywhere with no app to open. Same memory, same training — replies sound like your assistant, not a stranger's chatbot.

AI Reseller

Photograph books, electronics, LEGO, or shoes to get an identification, an estimated resale range, and a buy call — in the store aisle.

Daily Briefings

Weather, updates, and reminders that arrive on your schedule instead of waiting to be asked. You choose what it watches.

Company Directory

55,000 US company profiles, bulk-enriched offline by a locally-run LLM — the public surface and the baseline context every AI worker starts from.

Company Workspaces

Activate any company into a private workspace. Research runs in the background and the profile fills in — public intelligence and your private overlay kept distinctly labeled.

Knowledge Trainer

Upload documents, notes, and corrections once. Every employee inherits them, so answers carry your details and your way of doing things.

Writing & Email Assistants

Drafts blogs, follow-ups, and outreach in your tone, grounded in the documents you trained — not a blank page and not generic AI voice.

Architecture

How it's put together

Heynet is an orchestration layer over a shared memory and knowledge substrate. Each employee is a prompt-engineered persona with its own tools and channel; everything an employee learns writes back to one memory layer the whole staff reads from.

Channels

iPhone App

Native iOS + CallKit

Web App

SMS

Inbound Calls

Carrier forwarding → Twilio

Orchestration

Employee Router

Intent → right employee

Persona Layer

Mode + tone per user

Tool Layer

Calendar, search, web

Delivery Preferences

Push, SMS, in-app

AI Employees

Call Screener

Messenger

Reseller

Briefer

Writer & Email

Knowledge Trainer

Intelligence & Memory

Vector Memory

Calls, chats, documents

Knowledge Base

Your docs and corrections

Company Profiles

55,000 enriched records

Company Workspaces

Private overlay per company

Foundation

Hosted LLMs

GPT-4o, task-routed

Local LLM

Bulk profile enrichment

Python + Laravel

Voice/SMS backend + directory

MySQL + Pinecone

Twilio · FCM · Calendar

Perspective

How AI employees collaborate — at work and at home

The unlock isn't any single agent — it's shared state. A call comes in while you're in a meeting. The Call Screener answers, and because it can read your calendar it says so in its own words rather than reciting a canned greeting. The message gets transcribed, summarized in your assistant's voice, and texted to you — then written into chat history and vector memory. Next week, when the writing assistant drafts a follow-up, that call is context it already has. Nobody re-explained anything.

Collaboration happens through that memory layer, not through bots talking to each other. Every employee reads and writes the same store, so a handoff is just the next employee reading the file. It's what makes good human teams fast: shared context beats clever individuals.

Running both a personal product and a company product on one substrate taught me the sharper lesson. The consumer side proves the employee metaphor daily — a screener that blocks a robocall earns trust in a way a demo never does. The work side is where the metaphor has to hold up under scrutiny: an AI worker drafting to a customer needs governed permissions, cited sources, and an approval step. Same memory architecture, completely different contract with the user. Recognizing that they don't belong in one navigation was a product decision, not a technical one — Heynet Personal and Heynet Work now share infrastructure and nothing else.

Prompt engineering is the management layer throughout. Each employee has a persona spec — role, tone, boundaries, escalation rules — and designing those felt far less like writing prompts than like writing job descriptions.

Challenges

What was hard, and what I did about it

Knowing who a forwarded call is for

Every user forwards to one shared number, so identity rides on a carrier-dependent header that not every carrier sends. When it's missing, the flow falls back to asking the caller to key in a number — and the real lesson was instrumenting how often that fallback fires before designing around it.

A shared spam list is a shared liability

Flagging a number blocks it for everyone, which is powerful and dangerous in equal measure: one false positive on a pharmacy or a school is a support fire. Moving classification server-side means it now needs per-user scoping and confidence thresholds before it can promote anything globally.

Enrichment you can actually cite

Generating 55,000 company profiles is easy; generating ones you'd stake a sales call on is not. Every profile carries source URLs and a confidence score, unverifiable fields come back empty instead of guessed, and the enrichment prompt treats fetched web content as untrusted data — a scraped page should never get to issue instructions.

Tech stack

Built with

Swift / iOSPython / FlaskLaravel / PHPMySQLPineconeOpenAI APILocal LLM inferenceTwilioFirebase FCMGoogle App EnginePrompt engineering

Outcomes

What changed

  • 55,000company profiles bulk-enriched by a local LLM — the baseline context every AI worker starts from
  • 2 productsPersonal and Work, on one account, one memory layer, and one prompt architecture
  • 24/7call coverage — every unknown caller screened, transcribed, summarized, and delivered

Future roadmap

Where it goes next

Realtime voice: replace record-and-summarize with a streaming conversational agent that talks to callers live

Server-side spam classification with per-user scoping and confidence thresholds before global promotion

Company Intelligence OS: workstreams, tasks, approvals, and governed AI workers with scoped permissions

Meeting prep that assembles a cited brief from company intelligence and calendar context

Team workspaces — invite colleagues into a shared company workspace with role-based access

Provenance and freshness on every profile: entity resolution, change detection, and conflict handling

Next project

ABGi Marketing Automation