
CTO & Co-Founder @enrich_labs San Francisco Bay Area Previously led ML/AI/Infra @Scale_AI @Uber
Joined February 2025
- Tweets413
- Following195
- Followers495
- Likes296
I went back and checked: TK’s removal at Uber followed 4 months of controversy, an independent investigation, and investor pressure. He even retained his board seat.
Jensen was out of both roles within days, and NVIDIA board ruled out his return.
How is this maximizing shareholder value?
BREAKING: Nvidia, $NVDA, ends the week down -23% in its biggest weekly decline since October 2008.
The stock erased -$905 billion in market cap this week and officially closed down -$2.88 trillion since October 6th.
No comment from the company's board or interim CEO amid today's collapse.
8
143
Yunfeng Bai retweeted
1/ I ran 123 black friday campaigns that drove $210m+ in revenue for 55+ brands including Meta Oculus/Portal, Jack Archer, Udemy, and more.
here are the 12 tips to run BF/CM to maximize return
Read on >>
It’s very interesting to look at my GitHub history. Feb 2026 is around the time I started to use Conductor. With practice on operating multiple agents at the same time, and new ways to remove myself as the bottleneck, things plateaued around Aug - Oct, and now it’s increasing again with proactive PRs. A few observations:
1. My capacity on operating parallel coding agents is capped around 8. Since Feb I’ve had 5 PRs open at the median, 8 at the 90th percentile, and almost never more than 10.
2. Context switch cost is real. In 2025 the median PR was merged 5 minutes after it was opened. Now it’s 36 minutes.
3. With parallel coding agents, it’s much easier to get through smaller PRs and not so much for bigger ones. PRs under 50 lines went up 19x a month, PRs over 1,000 lines only 3x.
4. Lines of change didn’t see as dramatic an increase as PRs though. I merge ~7x more PRs a month, but change only ~3.6x more lines.
5. More productivity gains need to come from proactive agent proposals.
p.s. Code = ts/tsx/js/py/sql/css; lockfiles, JSON and bulk imports excluded
2
4
157
If you're running a fleet of models: even if every model in prod sits on the cost/quality Pareto frontier, your mixture of them might not.
The chart shows the cost and model-caused alert rate for different model mixes based on our production data (each chat session stays on the same model).
Model mixtures have a frontier of their own. It's worth looking at where your traffic mix sits.
2
4
121
I tried to let Claude Code and Codex review and fix each other's PRs. It spiraled into nits and bloated PRs, and I couldn't prompt my way out.
That’s one of many learnings from a year of running coding agents. Commits went up 3.7x along the way. Improvements came from finding where I was the bottleneck and removing myself:
1. Managing many coding agents got much easier with ADEs like Conductor @conductor_build (btw congrats on the acquisition)
2. Then running dev servers in parallel became the bottleneck. Spotlight is nice but only one session can run at a time. Took some effort to fix the dev env so each agent boots its own stack on its own ports. Also had to share dependencies so thousands of worktree dev envs don't blow up disk space.
3. Then validation became the bottleneck. Playwright MCP made browser validation much easier. I gradually moved to Codex browser and computer use (still better than Claude Code here) driving my main browser and laptop end to end. They now also handle my misc ops tasks, like reading FullStory sessions, setting up accounts, and updating website configs.
4. Then correctness became the bottleneck. To improve a non-deterministic agent, I let the coding agent run an eval-driven loop: write evals and confirm the agent fails them -> fix -> evals pass.
5. Then figuring out what to fix became the bottleneck. Now auto research agents dig through aggregated customer signals (e.g. intention mismatches), propose PRs on their own with the above techniques, and merge the low-risk ones.
What I still can't hand off:
Figuring out what to build. Getting real world feedback, a sense of where tech and the industry are going, and getting everyone aligned still takes a lot of human involvement.
Architecture. Coding agents with frontier models still don't hold the full picture. Their changes are often scoped narrowly.
Review. On the PRs I still merge myself, I rarely read line-level logic anymore, but over-engineering and weird implementations still get through.
So agents own execution and verification. I own what to build, architecture, taste, and saying when it's done.
For some reason there are way more security scans now - I suspect might be due to popularity with tools like github.com/projectdiscovery/…
Everyone should consider some auditing to make sure their site is not penetrated. CC or Codex can easily spot these attacks and make fixes.
3
87
We built ACO because personal agents buy differently from humans.
If you want to know why Muse or Instinct drops off on your store, this tracks it and Helena deploys the fixes.
Worth a spin if you care about agentic commerce.
Today, we're introducing Agent Conversion Optimization → agentconversion.ai
Muse, Instinct, Grok Bot already process $10B+ a year in GMV via agentic commerce. No GEO tool today tracks who they recommend.
So we built it.
Each agent behaves very differently from typical LLMs, so we track them separately, then Helena deploys the fixes.
1/ Daily tracking of who Muse, Instinct, Dots recommend, and why
2/ Share of recommendations vs your competitors
3/ Agent path to check out audit
4/ Fixes deployed by Helena to improve your discoverability
This is hard to do reliably with little bias given how these agents behave, so we're starting with only 100 initial customers.
Demo:
2
54
Yunfeng Bai retweeted
Today, Learnings is live in Helena.
Helena already reviews every channel weekly and turns your results into strategy. Now you can see it and steer it.
1/ See what actually worked on your brand. Your wins and kills on each channel, written as rules, not a dump of charts.
2/ Correct her once. Add "no hashtags" or "draft only" as a directive, and every draft, autopilot run and scheduled job follows it, on every channel.
3/ Pick the next test. Open the tab, choose a proposed experiment, ship it, and promote the winner to a validated finding.
Give it a try → hirehelena.com
Try this in ChatGPT after connecting Helena:
"Look at my last 30 days of sales and tell me which product I should be running ads on, then draft 3 ad variants for it."
Marketing team in a chat window. Link below 👇
its' official - we're live on ChatGPT plugins now!
give your chat the marketing superpower via the curated marketing skills/tools/workflows with our Helena agent - built upon decades of industry best practices and expertise scaling real businesses
available on ChatGPT plugins now👇:
chatgpt.com/plugins/plugin_a…
3
73
More buying decisions are starting with AI, yet most brands have no idea whether they’re even being recommended.
Now Helena captures GEO signals and fixes them. Give her a try and start optimizing AI discovery today!
Today, Helena handles your GEO.
Buyers ask ChatGPT what to buy before they ever reach your site. If you’re not in the answer, you never see that lost deal.
Helena asks ChatGPT, Perplexity and Gemini the questions your buyers ask, and shows you exactly where you show up and where a competitor does instead.
1/ Audit. Give Helena your brand, your site and up to 5 competitors. She runs your buyers’ questions across ChatGPT, Perplexity and Gemini and returns your
mention rate, citation rate and share of voice on each platform.
2/ Find the gatekeepers. See which articles AI keeps citing for your category. Those are the pages worth getting featured in.
3/ Catch what AI gets wrong. Stale pricing, missing differentiators, or a competitor named better at what you do best.
4/ Fix it, fastest first. Every gap maps to a ranked action list, from unblocking AI crawlers to on-site content, PR pitches, directories and reviews.
5/ Prove it worked. Helena re-runs the same questions 1 to 2 weeks later so you can compare before and after on each platform.
Give it a try today → hirehelena.com
2
63
Yunfeng Bai retweeted
Helena now captures leads for you → hirehelena.com
Give her one lead job and she finds the accounts, enriches the decision-makers and drafts the outreach in your voice. Nothing sends until you approve it.
1/ Outbound. "Find 50 ICP accounts in SaaS." Helena returns 50 accounts, each with an ICP score.
2/ Pipeline. Decision-makers enriched with emails and titles, pulled from Apollo.
3/ Campaigns. 36 outreach drafts in your voice, held until you hit send.
4/ Your whole stack. Email, social and ads in the same brand voice, through one approval gate.
Give it a try → hirehelena.com
If you're handling B2C, give this a read
and save it for future reference
if you learn social distribution, you will never beg for customers again.
you only need 2 things:
1. a product people actually want
2. a system that puts it in front of buyers every day
i've driven 200M+ views on ig + tiktok doing this for b2c.
the 10 steps:
5
49
if you learn social distribution, you will never beg for customers again.
you only need 2 things:
1. a product people actually want
2. a system that puts it in front of buyers every day
i've driven 200M+ views on ig + tiktok doing this for b2c.
the 10 steps:
My Mac kept going from 200 GB free space to zero in an afternoon. I was sure it was my hundreds of coding workspaces 🤔
Nope. It was the Aerial wallpaper.
macOS’s idleassetsd leaks half-finished video downloads into a root-only temp folder. 270 files, 750 MB each. du can’t even see them.
Check yours:
sudo du -sh /private/var/folders/zz/*/T/com.apple.idleassetsd
4
148
Helena can find anything on X for you
Use it to grow your account & engage with your community!
Same prompt to Claude and Helena: “Find me X posts to reply to.”
Claude can’t see X, so you get advice about where to look instead of posts.
Helena pulls the actual posts and drafts a reply for each one → hirehelena.com
1/ Real posts, not hypotheticals. Helena pulls the live posts you can reply to right now.
2/ A drafted reply under every post. Edit it, post it, move on to the next.
3/ No copy-pasting. You don’t paste posts into a chat window to get a response.
Give it a try → hirehelena.com
1
3
124
Yunfeng Bai retweeted
Same prompt to Claude and Helena: “Find me X posts to reply to.”
Claude can’t see X, so you get advice about where to look instead of posts.
Helena pulls the actual posts and drafts a reply for each one → hirehelena.com
1/ Real posts, not hypotheticals. Helena pulls the live posts you can reply to right now.
2/ A drafted reply under every post. Edit it, post it, move on to the next.
3/ No copy-pasting. You don’t paste posts into a chat window to get a response.
Give it a try → hirehelena.com
We've introduced pixel-perfect image editing to Helena!
Give it a try: hirehelena.com
Today, pixel-perfect image editing is live in Helena → hirehelena.com
The biggest challenge in marketing content generation is still consistency. Many image tools redraw the whole ad. The product shifts, the face changes, the logo drifts.
Helena's pixel-perfect image edititing copies every pixel you did not ask to change from the original. The approved asset stays the approved asset.
Some practical use cases include:
1/ Turn your exact product photo into an Instagram post:
The SKU, label, and packaging stay locked. Only the background, props, and caption layout change.
2/ Edit a Meta ad and keep the same character face:
Swap the hook, the outfit color, or the offer. The person legal approved is the person that runs.
3/ Localize one hero ad into Spanish, German, or Japanese:
The headline changes in place. The layout, logo, and product do not.
4/ Build a colorway set from one approved shot:
Red, black, limited edition. Same product, three ads, no reshoot and no off-brand drift.
5/ Cut that same file for Stories, feed, and email:
The crop changes. The brand does not.
Live now for all Helena users.
Give it a try → hirehelena.com
1
2
86
Yunfeng Bai retweeted
Today, pixel-perfect image editing is live in Helena → hirehelena.com
The biggest challenge in marketing content generation is still consistency. Many image tools redraw the whole ad. The product shifts, the face changes, the logo drifts.
Helena's pixel-perfect image edititing copies every pixel you did not ask to change from the original. The approved asset stays the approved asset.
Some practical use cases include:
1/ Turn your exact product photo into an Instagram post:
The SKU, label, and packaging stay locked. Only the background, props, and caption layout change.
2/ Edit a Meta ad and keep the same character face:
Swap the hook, the outfit color, or the offer. The person legal approved is the person that runs.
3/ Localize one hero ad into Spanish, German, or Japanese:
The headline changes in place. The layout, logo, and product do not.
4/ Build a colorway set from one approved shot:
Red, black, limited edition. Same product, three ads, no reshoot and no off-brand drift.
5/ Cut that same file for Stories, feed, and email:
The crop changes. The brand does not.
Live now for all Helena users.
Give it a try → hirehelena.com
















