Yunfeng Bai

Building Helena the marketing agent at Ex ML/AI/infra lead at Scale AI, Uber

[Yunfeng Bai @YunfengB]

The result: half the model bill of our single-model days, live quality trending up, and a verdict on every new model in days instead of a gamble.

Traffic share across the experiment: models earned share on evidence and lost it on repeating failures.

[Yunfeng Bai @YunfengB]

Every week there's a new model that looks better and cheaper on public benchmarks. Whether that’s true for your workload, only your own eval tells. Here’s the fleet we built to answer that continuously. https://t.co/BLEUmuRNSH

[Yunfeng Bai @YunfengB]

@SeijinJung Now it’s so much easier to manage recurring tasks!

[Seijin Jung @SeijinJung]

Introducing Task Feed for Helena.

It's your marketing command center.

One feed that pulls real-time intelligence from every platform you've connected: ads, email, content, seo. Turns it into the biggest growth moves across your channels.

A competitor launched new meta ads?
> Helena already flagged it with detailed analysis

Movements in your GEO and ai mentions?
> Flagged and prioritized

Negative keywords draining your google ad spend?
> Fix is staged and ready

She handles the monitoring, the analysis, the drafts.
You handle the last 1%: approve or reject in one click.

Live now on the right side of your dashboard.

p.s. the goal was always to make marketing feel like decision-making, not busywork. this gets us closer.

[Seijin Jung @SeijinJung]

a YC founder's linkedin impressions went up 3,074% in 2 weeks.

here's what happened 👇

he came to us with a common problem:
➤ inconsistent posting schedule, no traction
➤ no differentiated voice or insights
➤ writing content he found interesting vs. what his audience actually wanted

helena (our marketing agent) ran a full audit of his linkedin account, performance data, benchmarks, the works.

from there she built out a complete content pillar strategy for him to review. then she set up a recurring task to brainstorm topics for each pillar based on the latest trends, and drafted posts using linkedin best practices and our proprietary channel intelligence.

he reviewed and posted on a regular cadence.

results after 2 weeks:
➤ 2 posts went viral, 40k+ views in 3 days
➤ account-level impressions up 3,074%
➤ customer so happy he's offering to buy us meals

if your linkedin feels like it's going nowhere, it's probably not a content problem. it's a strategy problem.

[Yunfeng Bai @YunfengB]

@SeijinJung Great job Helena!

[Seijin Jung @SeijinJung]

Today, we're shipping our Marketing Intelligence Layer

120,000+ top Meta ads, emails, and growth playbooks, all curated for agents.

Every good marketer does the same thing before they write anything. They go study what's already working. That takes days (sometimes weeks).

We just made it take seconds.

Helena now reads from the largest curated marketing best practice repository ever built into an AI before she writes a single word for you.

→ 20k+ email templates from the best brands in dtc, saas, b2b
→ 100k+ Meta ad creatives with the hooks, copy, and angles that actually converted
→ brand briefs from award-winning campaigns
→ growth playbooks with real tactics and real numbers
→ brand identity references across thousands of companies

Live now for all Helena users today.

Try it → https://t.co/LbKm7d2Fjk

[Yunfeng Bai @YunfengB]

@SeijinJung Giving Helena a real library of what works makes a huge difference in making content!

[Yunfeng Bai @YunfengB]

I suddenly started to hear music from my laptop and it turns out Codex just being Chill 😂

[Seijin Jung @SeijinJung]

AI just ran its first billboard in Times Square.

Our AI marketer Helena did it. Solo.

Here’s exactly what she did:
→ read the brand and product guidelines end-to-end
→ researched current creative best practices
→ wrote the copy herself:
“this is AI slop. it still outperformed your ad agency.”
→ identified the 5 highest-visibility spots
→ cold-emailed the media vendor
→ got quoted $26k/day
→ negotiated it to $10k/day
→ sent the final deal for payment approval
→ launched it

No human picked the creative.
No human wrote the copy.
No human ran the negotiation.

She opened at $15k - Vendor countered $18k - She held at $12k - Settled at $10k; Methodical.

The copy is self-aware on purpose. Brand guidelines say be direct. The most direct thing an AI marketer can say is: yes this is AI-generated, and it still works.

One agent. Zero humans in the loop on the actual work.

Excited to see what Helena does next to push the boundaries of marketing.

[Yunfeng Bai @YunfengB]

@SeijinJung Helena has serious negotiation skills!

[Yunfeng Bai @YunfengB]

@SeijinJung Huge unlock for tech-forward marketing teams!

[Yunfeng Bai @YunfengB]

Helena MCP is a huge unlock for tech-forward marketing teams 🚀

[Seijin Jung @SeijinJung]

Today, we're shipping Helena MCP.

Connecting Claude to your ad account takes an afternoon with custom connectors.
Encoding 15 years of marketing pattern recognition into it takes something else.

That's what Helena is.

Connect your Claude to Helena, the most advanced AI marketer,
with 100+ custom skills and 3,000+ integrations,
purpose built by marketers who actually scaled $10M + businesses.

1 URL. 3 minutes.

→ Daily performance reviews across GA4, GSC, Meta Ads, Google Ads, social. automatically.
→ Writes SEO/GEO content that ranks. straight to WordPress, Webflow, or Framer.
→ Keyword research via GSC and Google Keyword Planner, weekly.
→ Launches and optimizes Meta and Google Ads from scratch.
→ Builds email campaigns in Klaviyo, Mailchimp, Brevo.
→ Sends a daily learning brief. what ran, what worked, what's being improved.

Real outcomes from early users:
→ SEO/GEO: B2B SaaS organic traffic up 85% in 6 weeks
→ Google Ads: 2x-ed a DTC brand's conversions in 3 weeks, scaled a health practice from $0 to $10k spend at 3.1 ROAS
→ Meta Ads: 200k+ purchase conversions for thousands of businesses
→ Email: $40k + in email revenue for hundreds of DTC brands

The setup is dead simple:
→ Go to Claude settings
→ Paste https://t.co/i4pY57mM0f
→ Authenticate your account
→ Done.

No CLI. No more random skill.md files. No more 47-tab morning routine.

Just your URL and 3 minutes.

[Yunfeng Bai @YunfengB]

🤔 From Artificial Analysis

[Yunfeng Bai @YunfengB]

Wow, Anthropic must be under tremendous competition pressure at mid-tier models wars:

> Sonnet 5 permanently at $2 per million input tokens and $10 per million output tokens

This is the pricing range (maybe even a bit lower) that competitors most aggressive at

[Yunfeng Bai @YunfengB]

@MichaelWaitze Nope, OpenAI started with in-house chip at least pre end of 2024

[Yunfeng Bai @YunfengB]

Anthropic docs such goldmine for context management: https://t.co/9zZABcgk9f
I think they understand agent context management better than any other LLM providers. These features like tool results clearing, thinking block clearing, mid conversation tool changes. All so related to actual agent usage.
Really hope other providers to catch up in the game.

[Yunfeng Bai @YunfengB]

Intelligently routing requests to different models - okay this seems great at first sight, but how do you deal with KV caching? I don't want to destroy input token cost by always refreshing cache. Is there a solution to this?

[Yunfeng Bai @YunfengB]

@MichaelWaitze No, I have been using Codex + GPT 5.6 Sol quite a bit recently and I can say there's not significant software edge from Anthropic at this point. Branding still huge though.

Custom chip - do you mean Cerebras with OpenAI? Since I don't think TPU really an "edge" right now.

[Yunfeng Bai @YunfengB]

@SeijinJung So excited to see how far Helena has come, and even more excited for what’s next 🚀🚀