AI Implementation
An end-to-end AI prospecting pipeline (17 stages)
I built a reusable AI “skill” that takes one company name and runs the whole prospecting flow — research across sources, a product-fit check, contact ranking, auto-enrich (email + phone), structured lead trackers (Notion plus auto-built, auto-formatted Google Sheets with linked Google Docs reports), and ready WhatsApp messages — with human approval before anything sends. The same flow also runs as a single command that executes about 17 stages and outputs a clean Vietnamese PDF report in 3–4 minutes.
Year :
2026
Industry :
B2B Sales / CRM
Client :
DAPLAST JSC
Project Duration :
2026
Problem :
Good prospecting is not one task; it is a chain of careful steps that must be done well every single time. Research the company. Judge whether it really fits our products. Find the right buyer, not just any contact. Get a correct email and phone. Check for compliance and trade signals. Log everything. Then write a first message that sounds human, not like spam. By hand, this takes a long time per lead, the quality changes with mood and energy, and it is almost impossible to keep consistent across hundreds of leads.
Solution :
I packaged the whole chain into one reusable AI skill, built and run on Claude.
You give it a company name, and it runs the full flow: research the company across several sources → check fit against our product range → find and rank the right contacts (the buyer first, then route contacts who can pass you to the buyer) → auto-get email and phone for the top few (via Apollo) → build a structured CRM in Notion (a Leads space → a page per company → a sub-page per contact) and/or a formatted Google Sheets tracker → draft natural, human-sounding WhatsApp messages → show them to me for approval → schedule the sends.
As a single command it runs about 17 stages automatically: read company context, search the web across multiple engines, deep-crawl the website (7 key pages), pull trade/shipment and registry data, find factories on maps, run compliance checks (sanctions, free-trade-agreement, and tariff), gather buying signals, enrich and verify contacts, score fit from 1 to 5, and render a clean Vietnamese PDF report — about 3–4 minutes per company, with results cached. Two rules are built in: data is cross-checked before it is saved, and the tone must pass a simple test — “could I send this same message to 10 different people in the same role and still feel right?”
Challenge :
Two parts were hard.
First, the outreach had to feel human and respectful — not scripted, not creepy, and without showing off that I had researched the person — so I wrote tone rules into the skill and kept the messages short and humble.
Second, control and safety: the AI never sends on its own. A person approves every message, and data is verified before it enters the CRM. The report also had to be genuinely useful — accurate compliance checks and a clean, readable Vietnamese PDF — so it could be shared with the team as-is. Getting that balance right — fast but safe, automatic but supervised — was the real work.
Summary :
This turned a task that used to take hours per company into a repeatable, controlled flow that anyone on the team can trust and reuse. It is a clean example of the main skill this job asks for: look at a real, messy workflow, find the parts that can be automated safely, and build an AI system around them — with a CRM, multichannel outreach, compliance checks, an auto-generated report, and a human in the loop. Used across 20+ export markets, it showed the flow holds up at real scale, not just in a demo.
More Projects
AI Implementation
An end-to-end AI prospecting pipeline (17 stages)
I built a reusable AI “skill” that takes one company name and runs the whole prospecting flow — research across sources, a product-fit check, contact ranking, auto-enrich (email + phone), structured lead trackers (Notion plus auto-built, auto-formatted Google Sheets with linked Google Docs reports), and ready WhatsApp messages — with human approval before anything sends. The same flow also runs as a single command that executes about 17 stages and outputs a clean Vietnamese PDF report in 3–4 minutes.
Year :
2026
Industry :
B2B Sales / CRM
Client :
DAPLAST JSC
Project Duration :
2026
Problem :
Good prospecting is not one task; it is a chain of careful steps that must be done well every single time. Research the company. Judge whether it really fits our products. Find the right buyer, not just any contact. Get a correct email and phone. Check for compliance and trade signals. Log everything. Then write a first message that sounds human, not like spam. By hand, this takes a long time per lead, the quality changes with mood and energy, and it is almost impossible to keep consistent across hundreds of leads.
Solution :
I packaged the whole chain into one reusable AI skill, built and run on Claude.
You give it a company name, and it runs the full flow: research the company across several sources → check fit against our product range → find and rank the right contacts (the buyer first, then route contacts who can pass you to the buyer) → auto-get email and phone for the top few (via Apollo) → build a structured CRM in Notion (a Leads space → a page per company → a sub-page per contact) and/or a formatted Google Sheets tracker → draft natural, human-sounding WhatsApp messages → show them to me for approval → schedule the sends.
As a single command it runs about 17 stages automatically: read company context, search the web across multiple engines, deep-crawl the website (7 key pages), pull trade/shipment and registry data, find factories on maps, run compliance checks (sanctions, free-trade-agreement, and tariff), gather buying signals, enrich and verify contacts, score fit from 1 to 5, and render a clean Vietnamese PDF report — about 3–4 minutes per company, with results cached. Two rules are built in: data is cross-checked before it is saved, and the tone must pass a simple test — “could I send this same message to 10 different people in the same role and still feel right?”
Challenge :
Two parts were hard.
First, the outreach had to feel human and respectful — not scripted, not creepy, and without showing off that I had researched the person — so I wrote tone rules into the skill and kept the messages short and humble.
Second, control and safety: the AI never sends on its own. A person approves every message, and data is verified before it enters the CRM. The report also had to be genuinely useful — accurate compliance checks and a clean, readable Vietnamese PDF — so it could be shared with the team as-is. Getting that balance right — fast but safe, automatic but supervised — was the real work.
Summary :
This turned a task that used to take hours per company into a repeatable, controlled flow that anyone on the team can trust and reuse. It is a clean example of the main skill this job asks for: look at a real, messy workflow, find the parts that can be automated safely, and build an AI system around them — with a CRM, multichannel outreach, compliance checks, an auto-generated report, and a human in the loop. Used across 20+ export markets, it showed the flow holds up at real scale, not just in a demo.
More Projects
AI Implementation
An end-to-end AI prospecting pipeline (17 stages)
I built a reusable AI “skill” that takes one company name and runs the whole prospecting flow — research across sources, a product-fit check, contact ranking, auto-enrich (email + phone), structured lead trackers (Notion plus auto-built, auto-formatted Google Sheets with linked Google Docs reports), and ready WhatsApp messages — with human approval before anything sends. The same flow also runs as a single command that executes about 17 stages and outputs a clean Vietnamese PDF report in 3–4 minutes.
Year :
2026
Industry :
B2B Sales / CRM
Client :
DAPLAST JSC
Project Duration :
2026
Problem :
Good prospecting is not one task; it is a chain of careful steps that must be done well every single time. Research the company. Judge whether it really fits our products. Find the right buyer, not just any contact. Get a correct email and phone. Check for compliance and trade signals. Log everything. Then write a first message that sounds human, not like spam. By hand, this takes a long time per lead, the quality changes with mood and energy, and it is almost impossible to keep consistent across hundreds of leads.
Solution :
I packaged the whole chain into one reusable AI skill, built and run on Claude.
You give it a company name, and it runs the full flow: research the company across several sources → check fit against our product range → find and rank the right contacts (the buyer first, then route contacts who can pass you to the buyer) → auto-get email and phone for the top few (via Apollo) → build a structured CRM in Notion (a Leads space → a page per company → a sub-page per contact) and/or a formatted Google Sheets tracker → draft natural, human-sounding WhatsApp messages → show them to me for approval → schedule the sends.
As a single command it runs about 17 stages automatically: read company context, search the web across multiple engines, deep-crawl the website (7 key pages), pull trade/shipment and registry data, find factories on maps, run compliance checks (sanctions, free-trade-agreement, and tariff), gather buying signals, enrich and verify contacts, score fit from 1 to 5, and render a clean Vietnamese PDF report — about 3–4 minutes per company, with results cached. Two rules are built in: data is cross-checked before it is saved, and the tone must pass a simple test — “could I send this same message to 10 different people in the same role and still feel right?”
Challenge :
Two parts were hard.
First, the outreach had to feel human and respectful — not scripted, not creepy, and without showing off that I had researched the person — so I wrote tone rules into the skill and kept the messages short and humble.
Second, control and safety: the AI never sends on its own. A person approves every message, and data is verified before it enters the CRM. The report also had to be genuinely useful — accurate compliance checks and a clean, readable Vietnamese PDF — so it could be shared with the team as-is. Getting that balance right — fast but safe, automatic but supervised — was the real work.
Summary :
This turned a task that used to take hours per company into a repeatable, controlled flow that anyone on the team can trust and reuse. It is a clean example of the main skill this job asks for: look at a real, messy workflow, find the parts that can be automated safely, and build an AI system around them — with a CRM, multichannel outreach, compliance checks, an auto-generated report, and a human in the loop. Used across 20+ export markets, it showed the flow holds up at real scale, not just in a demo.