Volodymyr Panchenko

Built DMarket. Exited. Now building Portal AI. Dad. 🇺🇦

[Volodymyr Panchenko @Portall]

0 marketing 🫡

[Volodymyr Panchenko @Portall]

@garrytan looks like this https://t.co/sMGkLsZpWX

[Hamudi @hamudinaanaa]

We went from 92% to 18% AI hallucinations.

we measured this on 30,000 personal AI agents. ask your AI about something you told it more than two weeks ago and it gets it right 8% of the time. the other 92% it makes something up.

force it to check its notes first? 82% success.

your AI doesn't have a hallucination problem. it has a context laziness problem. the answer is already in its files. it just doesn't look.

we fixed it with one sub-agent that runs before every reply. forces the lookup before generation starts. costs $0.04 per thousand queries. the laziness disappears structurally.

the full architecture: "do agents dream of electric sheep?" - link in comments

[Hamudi @hamudinaanaa]

proactive agent build day at @agihouse_org this sunday apr 26! come build with us.

the room: Rich Miner @richminer (Android co-founder, GV (Google Ventures) alum), Will Wang (CEO Even Realities @EvenRealities - smart glasses), Tobin South @TobinSouth (MCP & agents at Anthropic + Stanford), Melissa Pan @melissapan (Berkeley SKY, ex-Google).

four tracks: ambient agents that act before you ask. persistent memory that learns across sessions. agents for accessibility. best integration with Even Realities G2 smart glasses - real hardware, in the room, day one.

every track winner: $1,000 + G2 glasses for the whole team.

it's been a wild few weeks - so much excitement about our paper, our agents, our traction and fundraise. if you've been wanting to connect with me in person, this is it. apply on the page and come this sunday, i’m cohosting!

event: https://t.co/K3pZz87faP

[Volodymyr Panchenko @Portall]

@MarieNovosad13 @algozeus_ And you help us cater beter product every day 🙏

[Volodymyr Panchenko @Portall]

Day 1: 2 users.
Day 30: 20,000 agents. 33 countries.

63,000 files created. 306,000 Python scripts. 111,000 images. 2,367 Solidity smart contracts. 9 Fortran files.

$0 in marketing.

Imagine what happens when it's everyone.

https://t.co/YObH1jNQDz

[Volodymyr Panchenko @Portall]

@algozeus_ all of it. first few nudreds were from all my contacts in whatsapp and telegram. I just sent all of them at once the link with no explanation)

[Volodymyr Panchenko @Portall]

If you know someone who's already solving the chips + electricity + distribution equation - connect us.

Or DM me.

[Volodymyr Panchenko @Portall]

Day 19. 6,000 people. 6.5B tokens/day.

Someone launched a paid product without writing a line of code.
Someone landed their first client in 48 hours.
Someone won a six-figure government grant.
Someone's agent manages a multi-million dollar Shopify business daily.

They all 'just talked to AI.'

[Volodymyr Panchenko @Portall]

Our co-founder spent one weekend building a full game remastering pipeline.

Warcraft assets that took Blizzard teams months to modernize — done in hours with 200+ AI models orchestrated through Portal.

One person. One weekend. Thousands of beloved games could be next.

[Volodymyr Panchenko @Portall]

Google just dropped TurboQuant — compresses the AI memory cheat sheet from 32 bits to 3.

6x less memory. 8x faster. Zero accuracy loss.

We run 20,000 agents. 49% of our compute is just remembering.

The cost of an AI knowing you just got 6x cheaper.

$13/month per agent → $3 is now a 6-12 month prediction.

[Volodymyr Panchenko @Portall]

We swapped Opus 4.6, Gemini 3.1 Pro, and Sonnet 4.6 across 10% of users for 7 days.

Zero complaints.

Margin went from 58% to 92%.

If users can't tell — what are they paying for? Not the model. The memory, the orchestration, the relationship.

[Volodymyr Panchenko @Portall]

The most expensive part of running a persistent AI agent?

Not thinking. Remembering.

49% of our compute goes to rebuilding context at session start.

Messages 2 through 50 are nearly free (97% cache hits).

If you're pricing by messages, you're measuring the wrong thing.

[Volodymyr Panchenko @Portall]

At $49/month, only 47% of our users were organically profitable.

But when an 'unlimited' user spends $420 on compute and wins a $200K government grant — how do you price that?

Compute drops 50% every 6 months. Our margin goes from 58% to 89% in 12 months without changing prices.

The only race is distribution.

[Volodymyr Panchenko @Portall]

Our AI agents now have their own email addresses.

Name, memory, files, scheduled tasks, and now an inbox.

At what point do you stop calling it a chatbot?

[Volodymyr Panchenko @Portall]

Give real people an AI that remembers them, has a browser, writes files, sends emails, and runs on a schedule — and they don't use it to save time.

They use it to become someone they didn't know they could be.

That surprised me most.

[Volodymyr Panchenko @Portall]

20,062 isolated AI agents behind one Telegram webhook.

Each gets its own sandbox, browser, files, cron, and email.

Moved from 1 server to 40 shards in 65 seconds. Zero downtime.

One monkey patch saved $170K/month.

$13/agent to change someone's life.

[Volodymyr Panchenko @Portall]

What 20,000 people did with their AI in 30 days:

→ First SaaS built by someone who's never coded
→ $300K grant submitted at 11pm, resignation letter sent the next morning
→ Trading bot from simulation to live money in 24 hours
→ 28 English lessons rebuilt for a student with dyslexia

Life doesn't pause for clean narratives. Neither did their agent.

[Volodymyr Panchenko @Portall]

I was the other user on day 1.

30 days later — 2,062. 55% come back every day.

20,000 personal AI agents. 33 countries. $0 in marketing.

Mom texted she's in awe.

https://t.co/YObH1jNQDz