Case Study: Launched April 16, 2026

GALA — Autonomous B2B Lead Generation Agent

An autonomous agent that fills contact forms on websites for B2B lead generation. RoBERTa classifier determines strategy before opening the page. Result: peak of 85 inbound calls per day.

Andrey Rogovsky
12 min read· Updated

85

Calls/day (peak)

90

Active users

50-100ms

CPU inference

20K

Website dataset

99.8%

Ukraine (geo)

GALA is an autonomous agent that fills contact forms on websites for B2B lead generation. Technical core: RoBERTa classifier (PyTorch) trained on 20,000 real websites. Before the agent opens a page, the model determines strategy: simple / CSRF / complex. Result: 15–25% browser load savings, peak of 85 inbound calls per day.

Before April 16 — zero calls. Three and a half weeks of silence on Zadarma call tracking. After GALA launch on April 16: within 2–4 days — peak of 85 inbound calls per day. Stable baseline — 30–65 calls/day.

GALA is an autonomous agent that fills contact forms on websites for B2B lead generation. Finds potential clients, sends personalized messages, passes results to client CRM. Works 24×7 without human intervention. Technical core: RoBERTa classifier (PyTorch) trained on 20,000 real websites. Model determines form filling strategy before the browser opens the page. This gives 15–25% browser load savings compared to naive approach. CPU inference takes 50–100ms without GPU in production. Agent works autonomously, without manual intervention, generates live leads through contact forms on potential client websites.

Main Result#

These calls are not traffic to the GALA website. These are companies that received a message from the agent and are calling back clients. Call tracking captures feedback — direct confirmation that the agent generates live leads. Before GALA launch on April 16 — three and a half weeks of silence on Zadarma. After launch — within 2–4 days peak of 85 inbound calls per day. Stable baseline — 30–65 calls/day. Call geo: 99.8% Ukraine, mobile Intertelecom. This confirms that the agent generates live leads in the Ukrainian market, not just sending forms into the void.

85

Calls/day (peak)

30-65

Calls/day (baseline)

99.8%

Ukraine (geo)

2-4

Days to peak

GA4 Data (Jan 29 – Apr 28, 2026)#

90 active users over 90 days. 91 new users — almost all new. Average engagement time: 28 seconds. 493 events total.

Traffic Sources (sessions)

1

google / cpc

56 (51%) — paid traffic

2

direct

48 (44%) — direct visits, word of mouth

3

Teams CDN

2 — corporate Microsoft Teams — real B2B signal

Geo — Who Are GALA Clients

Singapore (13 users), Lviv (8), Kyiv (6), Wiesbaden DE (5), Amsterdam (3), Kharkiv (3). These are not foreigners or bots. These are Ukrainian entrepreneurs and diaspora running businesses abroad after 2022.

Pages

Home (GALA/Imaginify)

11 users32 eventsbounce: 62.5%

Landing "Gala | Lead Generation"

3 usersbounce: 0%

those who reached the product page read to the end

Why 0% Retention Is a Sign of the Right Product#

GA4 shows zero return rate. This is not a problem — it is confirmation of the model. GALA works on a set-and-forget principle: user logs in once, uploads a list of sites, configures message format — and leaves. The agent works autonomously, results flow into client CRM. This is the same model as Zapier or any background automation tool: real usage happens outside the browser. User does not return to the site because the product works without them. No reason to return — agent works 24×7 autonomously.

The right retention metrics for GALA are not session return rate in GA4, but agent activity: number of forms sent, volume of inbound calls on Zadarma, balance top-ups.

Unit Economics#

Average lead generation manager salary in Ukraine — ~$690/month. For $690 GALA sends 6,900 messages. At 10% conversion — 690 leads per month. A manager for the same budget can physically handle a maximum of ~200 contacts. Even if manager spends 5 minutes per site and works without breaks, they cannot compete with autonomous agent. GALA works 24×7 without sick leave, vacations, or fatigue. Scales without additional personnel costs. CRM integration available in Enterprise tier for automatic lead flow.

ManagerGALA
Cost/month$690$690
Leads/month~200~690
Availability8 hrs/day24×7
Sick leave / vacationYesNo
CRM integrationYesExtra fee

Technical Advantage: Why RoBERTa, Not GPT#

Not prompt engineering on top of GPT-4. Own trained model on real dataset with measurable effect. Classifier decides strategy before the browser opens the page. RoBERTa classifier trained on 20,000 real websites. Model analyzes HTML structure, determines form type, and passes strategy to browser agent. Agent executes only necessary steps for successful form submission. No extra requests, no extra load. CPU inference takes 50–100ms without GPU in production. This scales without growing token costs, unlike GPT-4.

simple

standard form, direct submit

CSRF

form with protection token, requires prior session

complex

multi-section form, file uploads, complex logic

Routing result: browser agent does not waste resources on unnecessary steps. Hence 15–25% load savings and 50–100ms CPU inference — no GPU in production.

Monetization#

1

Start

$0.10 per successful submission

2

Pro

$0.10 per form section (multi-section forms)

3

Enterprise

complex scenarios, CRM integration, call tracking — custom

What's Next#

  • SEO: organic coverage of queries "B2B lead generation", "cold outreach automation", "contact form filling"
  • International market: Singapore, Wiesbaden, Amsterdam in GA4 — testing EU audience with the same pain points
  • Dataset expansion: more form types, more languages for classifier
  • Retention through product: agent activity dashboard, submission stats, weekly digest in CRM

Stack

Python

Python

Primary language

PyTorch

PyTorch

ML framework for RoBERTa

RoBERTa

RoBERTa

Form strategy classifier

Playwright

Playwright

Browser automation

Zadarma

Zadarma

Call tracking

Astro

Astro

Frontend framework

Google Ads

Google Ads

Paid traffic

FAQ#

Why RoBERTa, not GPT?

Own trained model on 20,000 real websites gives measurable effect: 15–25% browser load savings. GPT-4 costs tokens per request, RoBERTa — 50–100ms CPU inference without GPU. This scales without growing token costs. Classifier decides strategy before the browser opens the page. Three classes: simple (direct submit), CSRF (requires prior session), complex (multi-section form). Routing result: browser agent does not waste resources on unnecessary steps. Not prompt engineering on top of GPT-4, but own trained model on real dataset with measurable effect. CPU inference without GPU in production gives infrastructure cost savings. Model trained on real data. Efficient and fast. Scalable.

How does the classifier work?

Model determines form filling strategy BEFORE the browser opens the page. Three classes: simple (direct submit), CSRF (requires prior session), complex (multi-section form). Result: agent does not waste resources on unnecessary steps. RoBERTa classifier trained on 20,000 real websites. CPU inference takes 50–100ms without GPU. This gives 15–25% browser load savings compared to naive approach. Model analyzes HTML structure, determines form type, and passes strategy to browser agent. Agent executes only necessary steps for successful form submission. No extra requests, no extra load. Classifier works fast and accurately. Training on real data gives high accuracy. Proven and reliable. Scalable efficiently.

Why is 0% retention good?

GALA works on a set-and-forget principle. User configures once and leaves. Agent works autonomously, results flow into CRM. Right metrics: number of forms sent, inbound calls, balance top-ups — not session return rate. This is the same model as Zapier or any background automation tool: real usage happens outside the browser. GA4 shows zero return rate, but this is not a problem — it is confirmation of the model. User logs in once, uploads a list of sites, configures message format — and leaves. Agent works autonomously 24×7 without human intervention. No reason to return to the site because the product works without it.

How much does it cost to use?

Start: $0.10 per successful submission. Pro: $0.10 per form section. Enterprise: custom with CRM integration and call tracking. Average lead generation manager salary in Ukraine — approximately $690 per month. For $690 GALA sends 6,900 messages. At 10% conversion — 690 leads per month. A manager for the same budget can physically handle a maximum of approximately 200 contacts — even if spending 5 minutes per site and working without breaks. GALA works 24×7 without sick leave and vacations. CRM integration available in Enterprise tier. Call tracking via Zadarma captures feedback from companies. Unit economics confirmed by real data. Scalable.

Who is the target audience?

Ukrainian entrepreneurs and diaspora running businesses abroad after 2022. GA4 shows: Singapore, Wiesbaden, Amsterdam — typical geography of Ukrainian B2B today. These are not foreigners or bots. These are Ukrainian entrepreneurs and diaspora running businesses abroad after 2022. Singapore (13 users), Lviv (8), Kyiv (6), Wiesbaden DE (5), Amsterdam (3), Kharkiv (3). Target audience is physically distributed worldwide, but seeks tools for working with Ukrainian and international markets. GALA hits this segment precisely. Call tracking shows 99.8% calls from Ukraine, mobile Intertelecom. This confirms that the agent generates live leads in the Ukrainian market. Product solves real problem. Proven by data.

Resources#

Andrey Rogovsky

Andrey Rogovsky

Senior AI Engineer · GenAI · MLOps · Cloud

25 years of infrastructure. Now I build AI that survives production with MCP + RAG + K8S.

More about the author →
© 2026 Andrey Rogovsky. All rights reserved.|Privacy