red 🍒

MIT CS/Neuro. Cofounder Avenir AI (a16z speedrun)
Building AI for the $1T employee benefits market.
I write about startups & building in public. AMA!

[red 🍒 @rediscoding]

Reintroducing myself after a break! 👋🏻
MIT grad, cofounder & CTO of Avenir.
Building AI for employee benefits.
Currently:
• Backed by a16z speedrun 🚀
• Live w/ Fortune 500 clients 💼
• Moved to SF from Boston 🌉
I write about startups, AI, & building in public.
Ask anything!

[red 🍒 @rediscoding]

@ezyemtechbotal @ukuzma2 I'm only a masters🥹

[red 🍒 @rediscoding]

sf is unreal 🔥

met my cracked founder friend today:

• biomedical eng phd
• MIT entrepreneur in residence
• Draper uni alum
• patented biosensors for sleep, weight management & stress
• so many awards it won't fit one screenshot

link to his startup in comments 👇

[red 🍒 @rediscoding]

@pigg_thomas Congrats 🎉 good luck w that process

[red 🍒 @rediscoding]

Hey X fam 🤖

- Today I built an agent loop system so Avenir can produce continuously smarter outputs w/out human intervention, using neuroscience principles + memory

- besides that, I ate a yummy milk cake at Sana'a cafe 😋

What did you ship today?

#buildinpublic

[red 🍒 @rediscoding]

@KantartzisThom Ikr 😋😋😋

[red 🍒 @rediscoding]

@vladzima Whoa... What was in the container 😂

[red 🍒 @rediscoding]

yeah! the direction we’re exploring is more like an agentic loop

each sub-agent has a specialized role + its own memory/context, then higher-level agents coordinate them. i think of it more like modular systems operating at different layers of abstraction rather than asking one network to consciously hold everything at once.

a few things that seem to matter a lot:
→ persistent memory so an agent learns “how you like the work done” and you don’t correct it twice
→ specialized sub-agents with narrower context instead of dumping everything into one context window
→ critic/verification loops before the result moves up a layer
→ different modes for different intent: experts can specify → execute, while vague requests trigger more clarification
→ learning from edits over time, then eventually using aggregate feedback to improve the default behavior

so “smarter” to me is partly model intelligence, but increasingly architecture + memory + feedback loops.

[red 🍒 @rediscoding]

@nyneurotech @ukuzma2

[red 🍒 @rediscoding]

@alpaysh I shed a tear 🥹

[red 🍒 @rediscoding]

@echo_yash 🥹

[red 🍒 @rediscoding]

just moved to SF after graduating MIT

since landing here:
- raised from a16z speedrun
- hit 6-figure ARR
- 650k+ impressions on X

if you’re in the city hmu. let’s grab coffee ☕

[red 🍒 @rediscoding]

@hsahu09 Wow thanks for the suggestion :) will keep in mind next time I visit India!

[red 🍒 @rediscoding]

gm X fam 🌞 I was locked in this morning with Claude working on agent loops, but here's me walking to the office.

Give me coworking space recs 👇🏻

[red 🍒 @rediscoding]

@MLCatttt @tmhammer love this 🔥

[red 🍒 @rediscoding]

@pardocar2000 @ukuzma2 yes 🙂‍↕️🙌🏻

[red 🍒 @rediscoding]

@_kvnloo Yeah kinda or like an MVP

[red 🍒 @rediscoding]

MIT taught me to ask 3 questions before building any feature:

1/ Does a current customer already need this?

2/ Will it make the product simpler or more complex?

3/ Can we test it with a manual version first?

What product q's saved you from feature bloat?

#startups

[red 🍒 @rediscoding]

@_kvnloo Nope it's my childhood nickname :) my name is Ari. "R + E"

[red 🍒 @rediscoding]

@ukuzma2 corgi cafe vibes indeed 🔥

[red 🍒 @rediscoding]

@thekkhaus so true 😂😭

[red 🍒 @rediscoding]

@AlanXiao67 go: jo

[red 🍒 @rediscoding]

@AmbikaSukla Simple & effective framework!

[red 🍒 @rediscoding]

@whatshoood why 63%? 👀 very precise