Anuj Fulari

Lead AI Engineer @ OtelAI, Building AI Agents for Humans and Agents, IITB

[Anuj Fulari @AnujFulari]

Hi X 👋 I'm Anuj, Lead AI Engineer at OtelAI, building AI agents for hospitality, automating work across every part of a hotel.

I spend my days on agents, LLM tooling & real-world integrations, and my nights thinking about entrepreneurship.

AI devs & founders, let's connect 🤝

[Anuj Fulari @AnujFulari]

🌸

[𝑝𝑢𝑟𝑒 @purelivn]

never stop hoping

[Anuj Fulari @AnujFulari]

Yeah however lot of things are still not discovered. I think model architectures will see some changes going forward. But we will definitely find out middle ground where both hardware and software i.e. now LLM in this case can support natively and efficiently. Like instead of etching whole model on a single hardware, we can each different kernels on small small hardware and they can act as building blocks for new chips.

[Anuj Fulari @AnujFulari]

@StevBuilds On the way 😀

[Anuj Fulari @AnujFulari]

@Shraddhaship Not anyone can sell though

[Anuj Fulari @AnujFulari]

@b1rkh0ff Let's connect 🤗

[Anuj Fulari @AnujFulari]

@itsankitakar It's actually changing. Everyone is doing vibe coding. It's the matter of who will test and maintain the code better.

At the end of the day, speed matters for every business. And vibe coding is just like nitro booster. Use it correctly, you will win the race.

[Anuj Fulari @AnujFulari]

I believe that even if we create a harness , a harness built with model which is trained to work with that harness, will be difficult to beat.

[hayato @hayatomaruu]

Anthropic engineer:

“Anthropic hired me at $350K–$950K because I know how to build harnesses.

99% of people still hand-roll multi-agent systems and wonder why they break.

I build systems that run loops + sandbox tools + turn yesterday’s logs into new skills — same model, 10x more reliable agents that keep improving overnight.”

Opus 5.5 → Agents → Harness → Loops → Graphs

In a short workshop, he shows exactly how to build one from scratch.

step 1 → harness handles loops, context, and sandboxing

step 2 → brain and hands separated — 60% faster to first token

step 3 → runs server-side, resumes from the log after crashes

step 4 → failure is cheap — retry and replay instead of starting over

step 5 → yesterday’s logs become new memory — the system wakes up smarter

Anthropic calls this “dreaming”.

Most people spend weeks building this by hand.
You don’t have to.

Bookmark and watch it.

Then read the full Opus 5.5 harness engineering guide below ↓

[Anuj Fulari @AnujFulari]

@DeepakNesss Context management depends lot on harness you use. I personally think claude code has done very well job in that and many will agree that it is the best out there

[Anuj Fulari @AnujFulari]

@NathenFernandes Could you share a secret? Ofcourse after competition is over

[Anuj Fulari @AnujFulari]

Just in: Cortex, from ByteDance Seed

The problem it tackles is one anyone building agents runs into. You give the model good context (API docs, lessons from past runs, updated facts) and it performs much better. Take that context away and it’s back to square one. So we keep stuffing prompts, building retrieval, and hoping the right memory gets pulled at the right time.

Their idea is to “consolidate” that context into the weights, loosely inspired by how the brain moves memories from short-term to long-term storage.

The approach is neat:

The model generates its own likely future questions about the context.
A copy of the model with the context teaches a copy without it.
Only small FFN updates get trained, and they merge back into the normal layers.

No new architecture, no retrieval step at inference. You end up with a regular model that just knows the thing now.

Some numbers that stood out:

50 fact edits applied one after another on a 14B model: it kept ~89% of the context gain with ~99.8% of original ability intact. AlphaEdit managed ~23%.
On ALFWorld (agent tasks), the consolidated model actually did better than just prompting with the context.
Smaller updates worked better than bigger ones.
Freezing old weights wasn’t enough to protect old knowledge. They needed to replay old questions.

That last point is underrated. Freezing parameters doesn’t freeze behavior.

It’s not free, though. Every update still costs a training run, and it only works for knowledge you can actually write down as context. But for agents that should get better at a specific job over time instead of re-reading the same playbook every call, this feels like the right direction.

[Anuj Fulari @AnujFulari]

@SciNatureNews How to visualise probability? 🥲 Like electron cloud? Instead of just electrons

[Anuj Fulari @AnujFulari]

@thefusioncube Yeah and considering how fast model architectures are evolving, it will be difficult with directly embedding single model on h/w. It is just gonna outdated in a 2 3 months

[Anuj Fulari @AnujFulari]

@RemiFabreRobot @antoinepirrone Is it available in India?

[Anuj Fulari @AnujFulari]

@thefusioncube Yeah true

[Anuj Fulari @AnujFulari]

@thefusioncube It's going to get more crazy in AI

[Anuj Fulari @AnujFulari]

@thefusioncube It's from Taalas, recently acquired by AMD. They are embedding models at hardware level

[Anuj Fulari @AnujFulari]

Where can I buy one?

[Elizaveta Zabrodskaya @Elizaveta_ai]

Microduck performance 😂

After they stole the show at NVIDIA GTC Berlin Keynote

[Anuj Fulari @AnujFulari]

@blessedofcodes Pytorch