CRM · live site
teamgoldllc.de ↗
Bilingual EN/DE marketing site with correct hreflang, four JSON-LD schema types, non-blocking font loading.
25 systems. Still running.
Built, shipped and operated by one person. Every number on this page is a measurement, not a claim.
Anything that costs money stops and asks.
Routine work runs unattended: a library of reusable agent skills, and a scored policy that decides which model and how much reasoning each task gets. One gate sits in front of every action with a price on it.
- 20+purpose-built agent skills
- 8escalation signals in the routing policy
- 24/7daemon, watched by a second process
The one everything else runs on
The screen is the easy part.
A cold-calling CRM with a browser softphone, an outbound AI voice agent, and the scoring that decides who gets called. Scroll: it comes apart.
- 01What the operator sees. One trade at a time, with the measured runway of each lead pool.
- 02What decides who gets called. Two scoring layers, geometric mean, an explainable tier per lead.
- 03What can stop it. Spend cap, call window, and a kill switch no single toggle overrides.
- 100k+leads scored and routed
- 7rotating outbound numbers
- 14+A/B-tested script versions
Scoring layer
lead · hvac · reachable612S
lead · roofer · mailbox388A
lead · plumber · no site271C
Sample rows, scored and tiered by the real thresholds. No customer data appears on this page.
Guardrail layer
spend capcall windowkill switchrow-level security
One multiplication destroyed the ranking.
The first scoring version multiplied need by budget by reachability. Mathematically reasonable, and it pushed almost the whole dataset onto a handful of duplicate scores. Drag the sliders and watch both formulas at once.
Measured on the real dataset: 9,400 of 59,725 leads landed on the identical score 35. The bars are all 9,261 combinations of the three inputs in steps of 5, in 20 bins.
Old, raw product0tier
New, geometric mean0tier
Every possible lead, both formulas
score 01000
It has to be right at 3am, not just at noon.
An always-on pipeline that finds local businesses, checks they still exist, and reads their own websites for signal. Two scrapers cross-verify each other before anything is scored.
- 2independent scrapers, cross-verified
- 30+business signals per lead
Live · offset paging against a table that keeps changing
Both lanes read the same 11 rows, the same two writes land mid-scan, and only one lane gets it right. Press play. The general-purpose fix is on GitHub as postgrest-keyset-page.
OFFSET n LIMIT 30 duplicates · 0 skipped
WHERE id < last_seen LIMIT 30 duplicates · 0 skipped
Row 11 gets inserted, then row 6 gets deleted, both mid-scan. Exactly the timing that broke this in production.
Measured on the real dataset: a 20-minute scan with 43 concurrent inserts produced exactly 20 duplicate reads under offset paging.
A dead browser looks exactly like a healthy one.
The watchdog restarts the scraper only after confirming real CPU activity over a window. Checking whether the process id still exists proves nothing.
A null is not a zero.
"Never measured" and "measured, and it is zero" look identical in a spreadsheet and mean opposite things to a scoring model.
I measured my own security tool. It was at 80%.
A dependency-free pre-commit hook that stops API keys reaching a commit, written after five live credentials sat in a repo for months. Building a labelled corpus of 61 cases surfaced three bugs, all false negatives: the failure mode that looks exactly like success.
- 98tests, 98% coverage
- 11CI jobs, 3 operating systems by 3 Python versions
- 80 → 100%recall, after measuring it. Precision 91.4 → 100%
Live · paste something and watch
The real patterns, ported to JavaScript. Four lines are pre-filled and each produces a different verdict: a catch, a placeholder, the prose that caused a real false positive, and an explicit opt-out.
Runs entirely in this browser. Nothing you type into the scanner leaves the page. View the repository.
Things other people actually open.
Two of the twenty-five are public and running right now. Both are screenshots of the live thing, not mockups.
teamgoldllc.deBilingual EN and DE with correct hreflang, four JSON-LD schema types, non-blocking font loading.
Portrait, same buildMobile is the first target, not a pass at the end.
FahmA side project, phone-first: one HTML file, no build step, 500+ Quranic words with word-by-word breakdowns.
Missed Call Rescue · delivered, and honest about what is unproven
A missed call becomes a text to the caller, a logged lead, and an alert to the owner, in under 60 seconds. Built for a two-person crew who are on a roof when the phone rings.
Two servers, administered by hand.
A Windows box and an Ubuntu ARM instance. SSH password authentication off entirely, keys only. Debug ports bind to loopback. Secrets live in the OS keystore, enforced by the hook above and again in CI.
Nineteen more, to the same standard.
Smaller in scope than the six above. Each one a real, independently working subsystem, not a stub.
CRM · Data
Lead Scoring & ICP Engine
Two independent scoring layers: geometric-mean need/budget/reachability, and a separate ICP-fit score with its own confidence metric.
Data · AI
Multi-Source Data Fusion
Merges signals from several observation sources into one profile per contact; reports confidence separately from freshness.
Data · AI
Hiring-Signal Detection
Parses a business's own website for hiring activity, ad-spend evidence, tech stack and multi-location signals, from data already fetched.
Data · AI
Multi-Source Lead Enrichment
Five more signals out of one page load already being made, with 0 additional network calls, plus public licence-registry matching that needs two independent fields to confirm a name.
AI · CRM
Two-Channel Call Audio
Splits recorded calls into separate customer and agent channels before transcription; measurably better than letting the model guess the speaker.
AI · CRM
AI Call Review & Coaching
LLM-generated rep feedback gated behind human approve, edit or reject, plus a blind grading mode so review is not biased by the AI's verdict.
CRM
Multi-Channel Outreach
Coordinates an email warm-up ramp, a queued LinkedIn channel, and a layer that labels measured numbers apart from self-reported ones.
CRM · Data
Controlled Experiment Framework
Test-cohort system with independent locks keeping test leads out of production automation, plus a self-releasing batch mechanism.
AI · Security
Custom MCP Tooling
Model Context Protocol servers exposing read-only business data and a security-testing toolkit, scope-gated off by default.
AI
Human-in-the-Loop Control
Telegram approval bridge letting a long-running agent pause and ask before any money, access or direction decision.
AI
Model / Effort Routing Policy
Scored, versioned policy deciding which model and how much reasoning effort a task gets, across eight escalation signals.
Data · AI
Self-Healing Task Watchdog
A separate monitor checking whether the checker is still alive, built after a daemon died silently for three days.
Security
Authorized Testing Toolkit
Wraps nmap, nikto, sqlmap and whatweb behind a scope gate that refuses anything outside localhost and private ranges unless deliberately unlocked.
CRM
Compliance Call Windowing
A legal calling-hours check kept strictly separate from reachability scoring. The two answer different questions.
CRM · Data
Inbound Routing with a Hard Fallback
A rep pool rings first, with a server-enforced lock forcing a voicemail fallback regardless of any dashboard toggle.
AI · CRM
Weekly Niche-Focus Controller
Rotates which market segment the calling operation focuses on, driven by measured rep success rates.
AI · Personal
YouTube Research Pipeline
RSS-based channel monitoring with no paid API, feeding a pipeline that extracts business and engineering lessons.
Personal · public repo
Fahm, Arabic vocabulary
Single-file, build-free web app for Quranic Arabic: 500+ words, word-by-word breakdowns, custom vocabulary import.
Personal
Early projects
A chess engine, a crypto-trading backtester, and classic learn-to-code exercises. Kept private, and the actual starting point.
Built by directing coding agents deliberately: writing the specification, setting the guardrails, verifying the result. No degree; what replaced it was building real systems, operating them, and fixing what broke. No customer data, phone numbers or business metrics appear anywhere on this page or in the linked repositories.
Who builds this
Built by one person. Here is who.
How he works, and where he came from. The stack that shipped it all is right below.
How I work
I build fast by directing AI coding agents deliberately — writing the specification, setting the guardrails, and verifying the result — rather than typing every line by hand. That's a real, current skill: the systems above aren't demos, they run a real business every day, and treating "the agent said it's done" as a claim to verify rather than a fact was the single habit that mattered most in building them.
No customer data, phone numbers, or business metrics appear anywhere on this page or in the linked repositories; numbers are rounded, not exact.
Background
I didn't come to this through a computer science degree. The short version:
- ~19–21Ran online businesses — e-commerce, affiliate marketing, a YouTube channel. Learned how to get someone's attention and turn it into a customer, long before I could write a line of code.
- ~20Started learning to code, at first alongside the business work.
- 22Moved to Egypt and shifted focus fully to software.
- 22–23Freelance client work, then building my own acquisition infrastructure — cold calling, AI voice calling, email and outreach — for Team Gold, the business behind most of the systems on this page.
- 23Today: designing, building, and operating the systems above.
No university degree — what replaced it was building real systems, operating them, and fixing what broke.
Skills & stack
What actually shipped the systems above — not an aspirational list.
Tap any item for what it means in plain words — no jargon.
Languages
Application & Data
AI & Agents
Voice & Messaging
Infra & Ops
Ask, instead of reading all of it.
Pick one of the five, or type your own.