Andrei Sidorov

I explore crypto and build little tools with AI. Sharing my projects, things I learn, and what doesn't work. Open to freelance projects.

[Andrei Sidorov @madrew_st]

I've been in crypto since 2022, researching projects and working as a Web3 assistant.

Now I'm starting a new chapter: building useful tools with AI and sharing what I learn here.

Open to freelance projects, collaborations, and new roles. Ready to learn fast and give it my all.

[Andrei Sidorov @madrew_st]

A CRM with AI under the hood?

I've just finished an exciting project: a personal CRM system built from start to finish with vibe coding.

It all started with this setup:
/ @OpenAI - discussion and prompts
/ @claudeai - development

But… a couple of days, and the weekly limits were gone… so I finished the work entirely in Codex, and honestly, I'm not complaining.

Sure, there's a lot I don't understand, but I definitely wouldn't say it's unusable.

Now a bit about what the idea was and whether it all turned out the way I wanted.

The original request: a CRM with call transcription, AI analysis and preparation for investor calls.

At first we wanted to build it all in @NotionHQ, but then the question came up: why, if we could come up with our own?

It was purely my own wish to try building something with AI, so I set the task myself: let's do it, but make it cheaper than Notion.

So I looked for ways to keep maintenance within $24 a month, and ended up with this stack for a base of 200 people and 75 hours of recordings per month:

/ Analysis and preparation ~$2
Qwen via OpenRouter

/ Transcription ~$7.50
Soniox AI

/ Server — $5.76
HOSTKEY

That's ~$15.50 in total, or a saving of more than 35% a month.

We also plan to move transcription to Pocket, but we'll figure that out once we have it in hand.

For now, I'm waiting for comments from experts: what could have been done differently, or what else I could take on for practice :)

[Andrei Sidorov @madrew_st]

Vibe coding a personal CRM

For the past couple of days, I’ve been building a CRM to:

Transcribe calls → Analyze them → Help prepare for the next ones

Small scale: around 200 contacts, aiming for 300, and 75 hours of recordings per month.

Good practice with AI tools for me, and a way to help a friend with his negotiations.

Progress so far:
/ Local version working
/ Tech stack agreed
/ Integration groundwork in progress

Surprisingly, most of my time went into shaping the idea, the design, and how it all works. I’ll keep looking for ways to speed that up.

I think a couple more days should do it. Then I’ll share my overall thoughts.

[Andrei Sidorov @madrew_st]

Either you keep up, or you’re fucked!

That’s exactly how I want to start today’s post!

Today, I had a call with my mentor, who’s been encouraging me for a long time to take a course at the University of Artificial Intelligence and land an internship there!

And of course, I want to do it myself. But the course was created a year ago. In AI, that’s an incredibly long time. Everything changes by 540 degrees!

The information, the tools, and a lot of the other material in the course are outdated. That makes the assignments so difficult to get through that, even when you really want to do them, you end up thinking, “Fuck this…”

But somehow, it only occurred to me today that I could give Claude or Codex access to the course and ask it to rebuild it with updated material, in a way that actually makes sense!

So today, I burned through all my usage limits on this task. Tomorrow, I’ll keep going, because it’s not finished yet!

How is your week going?

[Andrei Sidorov @madrew_st]

Today I followed 50 orders through an n8n workflow and saw how data really moves from node to node.

Sample Order Data → Split Out → Filter Americas → Add 10% Tax → Total Price

The orders arrive as one array inside a single item. That matters, because n8n passes data between nodes as a list of items, and most nodes run once per item. One item with 50 orders inside is not the same as 50 items.

That's exactly what Split Out fixes. It turns the array into 50 separate items, so every next node handles each order on its own, and Item Linking keeps track of which output came from which input.

To check what came out, n8n shows the same output in three views: Schema for structure and types, Table for rows, JSON for raw values and paths. Looking at the types led to the next trap: 249 and "249" are not the same value. One is a number, the other a string, so before any math you convert it with {{ Number($json.total) }}.

Once the items are clean, the order of steps matters too. Filter Americas goes before Add 10% Tax. The result is the same, but the workflow doesn't add tax to orders it's about to drop.

Every step reads its data through expressions like {{ $json.total }}, and one quiz question showed the catch: insert a Filter before a Slack node, and $json in Slack now comes from the Filter's output. A named reference like $('Get Users') keeps the source explicit.

Section 2 of n8n Essentials (N8N101) continues in the next lesson.

Which data trap would you warn others about first?

@n8n_io automation

[Andrei Sidorov @madrew_st]

Chat or AI agent? Everyone's hyped about agents, but most people don't need one yet.

Here's the actual difference:

# Chat (ChatGPT, Claude, Gemini)

Open it and go. Best for questions, writing and ideas. Around $20/month or free. The catch: it only works while you're there, and the result stays inside the chat.

# Agent on your computer

Opens and edits your files, runs programs on its own. Great when your work lives on your device. But it stops when your computer is off, and you need to be careful with what you let it access.

# Agent on a VPS

Runs 24/7: monitors, automates, works while you sleep. But now you pay for the server and for the model via API, and long tasks can get expensive.

My rule:
→ simple request: chat
→ need to build something: agent
→ needs to run without you: agent on a VPS

To keep API costs down, I route different tasks to different models instead of using the most expensive one for everything.

What do you use, and for what?

[Andrei Sidorov @madrew_st]

How do you overload a workflow in @n8n_io ?

I built two workflows that do the exact same job and overloaded one of them on purpose. Result: 22 nodes vs 20 nodes. The count isn't the point. What matters is what each node actually does.

The task:
Text or voice in Telegram → a row in Google Sheets → a daily AI summary

Each version has two workflows: one receives the messages, the other builds the report.

INTAKE
• Overloaded: a text message passes through 4 nodes before it becomes a record
• Clean: one Code node builds the record for both text and voice, and the other nodes handle checks

REPORT
• Overloaded: read rows → filter by day → join texts → prepare notes → AI classify → summary
• Clean: one AI call and fewer in-between steps where things can break

BONUS
• A 5-node block in every workflow that creates a new sheet automatically when a new month starts

I didn't measure speed, cost or load, and that's on purpose. The goal was to put n8n Quickstart (QS101) into practice and see what one task looks like at different levels of complexity.

Takeaway: the longer the workflow, the more chances to make a mistake. And when something breaks, the harder it is to find and fix.

Which version would you pick, and what else would you simplify?

[Andrei Sidorov @madrew_st]

New challenge: I'm building the same n8n workflow twice.

One clean. One overloaded on purpose. Same job, then a side-by-side comparison.

I'm about 35% in, by my own estimate. The architecture is planned, and in the first version the voice and text branches already merge into one flow. Next come tests and the second build.

Guess which one I'm building first?

@n8n_io automation

[Andrei Sidorov @madrew_st]

It's too late to get into AI.

Almost everyone who wants to start in this field has had that thought. I had it too: for a long time I put off learning AI tools, sure the train had already left and there was no catching it. But is it actually true?

Once I realized I had nothing to lose, I started. And, as it usually goes, I suddenly saw people everywhere building big projects with AI, and companies saying AI agents already do half of their work...

It feels like everyone has already plugged AI in wherever they can. So where do you even fit in, when jobs themselves are on the line?

If that sounds like you, try one thing: zoom out. Ask random people on the street how they, or people they know, use AI.

You might be surprised: 90% will say they don't use it at all, or that the most they do is ask for a dinner idea based on what's in the fridge.

Why? Because we humans are pretty lazy. Most people won't even bother writing a prompt, and when the answer isn't what they expected, they say: "This AI thing is dumb!"

So my takeaway is simple: don't be lazy, just act. Even if it doesn't work out, you lose nothing. AI is useful for far more than work.

And worst case, if there's no room left in the AI bubble, or AI takes every job out there, remember: nails won't paint themselves ;)

What do you think about this?

[Andrei Sidorov @madrew_st]

Congratulate me on finishing my first n8n course! QS101 Quickstart is done, passing grade and certificate included.

The last block went beyond regular workflows: AI Agents. I put together a customer service agent with a chat trigger, a language model, memory and tools that let it reach customer and order data.

Two things I won't forget. Simple Memory only keeps recent context, it doesn't train the model. And the agent picks tools by their descriptions and instructions, so it can still pick the wrong one.

5 practice workflows in, 3 courses next: Essentials, Integrations and In Practice. First step done, more to come.

What would you build with an AI agent first?

@n8n automation

[Andrei Sidorov @madrew_st]

One focus again: automation. I worked through n8n's Quickstart lessons and condensed them into a 19-section cheat sheet covering 23 nodes.

The detail worth remembering: Merge has 5 modes, and Merge by Fields joins data like SQL, but the field names go in as plain text, not as expressions.

In the practice task I merged 30 orders with 10 customers by customerID, calculated each order total, and built a European CSV report plus a regional summary for Discord.

Error handling has its own toolkit: 3 On Error modes, retries up to 5 times, and a separate Error Workflow for production failures.

What node would you learn next?

@n8n_io automation

[Andrei Sidorov @madrew_st]

@Sudiarta1991 @0xAsta @jumperapp Now I understand, thank you! I have completed tasks, but I did it natively, using the bridge itself. There was no goal to earn more points. It’s just a really cool product.

[Andrei Sidorov @madrew_st]

Busy day, so I kept it to one track: automation. This time I went to the source - n8n's own course, not just GPT or Claude walkthroughs.

I finished 4 of the 5 parts of the Getting started section: the editor, a mini-workflow, a real-world use case and exporting workflows. The 15-question quiz is next.

The workflow I built runs every Monday at 9am or on demand. It pulls 30 orders from a warehouse API, checks their status and splits them: 5 are still processing, 25 are booked. The booked branch is counted in a Code node and handed to Discord.

Short on time, not on consistency.

What would you automate first with a weekly trigger?

@n8n_io

[Andrei Sidorov @madrew_st]

@0xAsta @jumperapp Believe in something?)

[Andrei Sidorov @madrew_st]

Share: how much XP and badges you have in @AbstractChain, friends)

Or maybe you’d rather share how much XP you have in @jumperapp?

Believe, XP = 1$ and GN

[Andrei Sidorov @madrew_st]

My n8n Telegram bot now turns a message like "Coffee 11 GEL" into a clean expense record.

I also modeled whether $100k in stablecoins can pay $1,000/month within 5 years: a constant 9% APY gets there in month 51, a 3.71% Aave snapshot needs ~421 months.

AUTOMATION · @n8n_io
Task: log daily expenses straight from Telegram and get a daily or monthly summary.

Message → access check → route (report / help / voice / expense) → AI extracts the fields → JavaScript check → Google Sheets → reply

✓ AI via @OpenRouter returns amount, currency, category and description in a fixed format
✓ A Code node validates that output before anything is saved
✓ Report commands read the sheet and total each currency separately
✓ Tested with "Coffee 11 GEL" and "Subscription 20 USD"

Open issues:
• "Subscription 20 USD" was parsed correctly but came back needs_review=true. The rule in the system prompt needs tightening.
• /today built the report, but TelegramReport failed binding chat_id via .item. Proposed fix: {{ $json.chat_id }} from BuildReport's own output, not confirmed yet.
• Passing the audio file for voice messages was hard to find in the new n8n interface.

WEB3 RESEARCH
Question: can $100,000 in stablecoins reach $1,000/month of passive income within 5 years? Income reinvested, no top-ups. Lending markets, liquidity pools and looping, only protocols with TVL above $50M and 2+ years of history.

Months to $1,000/month (model, scenario rates):
█ within 5 years · ░ misses · 1 block ≈ 4 months
████████ 33 · 10% APY, constant
█████████████ 51 · 9% APY, constant
██████████████ 54 · mixed positions + looping (capital ≈ $144,788)
░░░░░░░░░░░░░░░ not reached by month 60 · strict caps, ≈ $826.63/mo
░░░░░░░░░░░░░░░░░░… 421 · Aave snapshot ≈3.71% APY

Strict caps: looping 20%, pools 30%, lending 40%, reserve 10%.

How I worked:
• Observed data kept apart from model assumptions
• @DefiLlama for rates, TVL, market and available liquidity (@aave, @CurveFinance, @pendle_fi)
• Official Aave, @Morpho, Pendle and @ethena docs for collateral, debt, LTV/HF, liquidation, term, exit terms and base-asset risks
• Every data point saved with its date and link; model and sensitivity in a spreadsheet
• @OpenAI / @claudeai only to double-check the math, never as proof of yield

Verdict: on paper the goal needs about 9% average APY with reinvestment. A strategy that holds that for 5 years is not confirmed: rates move, and looping adds debt, liquidation and concentration risk. Learning model; no real funds or wallet involved.

What would you do next: fix the bot first, or stress-test the model further?

[Andrei Sidorov @madrew_st]

My @n8n_io Telegram bot now handles voice messages up to 30 s.

I also built a repeatable way to vet crypto projects: in a test @aave model, 12.10 USDC of interest became -$19.90 once fees and my time were counted.

AUTOMATION
Before: text only
Now: text + voice up to 30 s

Voice → transcribe → AI parse → check → Google Sheets → reply

The reply shows the transcript, and the sheet stores it. Too-long recordings and failed requests are handled.
Stack: n8n, Telegram, Google Sheets, OpenRouter

WEB3 RESEARCH
Use case → tech → risks → economics → dependencies → decision
Learning cases: USDC in Aave, DePIN

1,000 USDC, 90 days, 5% APY (test numbers)
████████████ +12.10 USDC interest
▏ $0.10 left after $12 fees
░░░░░░░░░░░░░░░░░░░░ -$19.90 once my time is counted

Lesson: fees and time can turn a "yield" into a loss.

What would you do next, with the bot or with the research?

[Andrei Sidorov @madrew_st]

Learning AI Automation & Web3 Research

Today:
- a Telegram bot in @n8n_io + @OpenRouter: sorts messages, sums them up, logs to Google Sheets
- an @aave liquidation model: HF 1.6, HF = 1 at a 37.5% WETH drop

Goal: go pro in both

[Andrei Sidorov @madrew_st]

Finally tried @claudeai after spotting a 100% cashback offer via my @Bybit_Official card. First task: troubleshooting my old laptop. It now feels faster and runs cooler!

What would you automate or build with AI? Drop ideas for my next project.