Jainit Purohit

CTO @ Terrabase, AI agents, eval harnesses, decision infrastructure
Meditation | Dhamma | phenomenology

[Jainit Purohit @mjainit]

May you find release with boundless freedom,
With formless expansion stretching in all directions,
Boundless, immeasurable, limitless, with goodwill in its core, and
filled with joy, reaching far, evermore.

[Jainit Purohit @mjainit]

🚨 Hiring Alert - Terrabase is hiring 3 senior engineers.

We're building a self-learning context layer and AI analytics harness at enterprise scale. Tough problems to solve, fast learning, and staying at the forefront of how AI systems actually work in production.

Work directly with Fortune 500 clients. Build context layers and AI agents that hold accuracy at scale. Ship systems that compound in value with every run.

If that sounds like where you want to be, apply below.

3 roles. All Remote.

Platform: https://t.co/LEnp6uVhf0
AI Agents & Evals: https://t.co/8Uasx7aam0
Applied ML: https://t.co/4HcIvVESj3

#Hiring #StartupJobs #EarlyStageStartup #FutureUnicorn #JoinOurTeam #BuildWithUs #CareerOpportunity #AI

[Soumya Jain @wild_and_empty]

One of the best guided meditations out there 🥹:

https://t.co/fHhSNYwMkP

[François Chollet @fchollet]

Agentic coding is a form of machine learning. Generated code is best treated as a blackbox artifact whose behavior and generalization should be managed via empirical evaluation, like with any ML model.

[Soumya Jain @wild_and_empty]

building AI systems, studying human systems, debugging my own.

unfortunately all three have context problems.

[kache @yacineMTB]

you can outsource your thinking
but you cannot outsource your understanding

[Jainit Purohit @mjainit]

@thsottiaux Codex has been great, thanks! One issue though on Mac: in long threads, scrolling up sometimes jumps to random positions much higher in the thread. It breaks context and you have to scroll back down again. Pretty frustrating, would be great to see this fixed.

[Harrison Chase @hwchase17]

https://t.co/aZnKZ1SVXB

[Soumya Jain @wild_and_empty]

This is exactly why trace data shouldn’t just sit in an observability bucket.

One layer below this is -- reviewed traces can also teach the agent how to work better, not just tell us whether the final answer was good. You can look at strong runs, review the workflow itself, tool calls, handoffs, context gathering, etc and turn that into a retrievable execution layer.

So not just a data asset for analysis, but a learning layer for better runtime behavior.

[Jainit Purohit @mjainit]

@manthanguptaa same fear cycle every year.
first "software engineering is solved", now "agents running companies".
wonder how many actually survive production and real users
most teams still fighting evals and edge cases. every line shipped is future maintenance debt.

[Jainit Purohit @mjainit]

@pmarca You’re conflating rumination with introspection.

Rumination reinforces negative pathways.
Introspection enables metacognition and error correction.

No cognitive tool is inherently good or bad. Outcomes depend on whether it produces emotional loops or better models of reality.

[Jainit Purohit @mjainit]

@elonmusk You’re conflating rumination with introspection.

Rumination reinforces negative pathways.
Introspection enables metacognition and error correction.

No cognitive tool is inherently good or bad. Outcomes depend on whether it produces emotional loops or better models of reality.

[Jainit Purohit @mjainit]

yep, this is basically the exact workflow I use right now.

I intended to automate the entire self-improvement loop but apart from obvious overfitting issues, you start seeing unknown behaviors and unnecessary layers of abstraction creeping in. The last remaining stabilizing step I had to add was a human-in-the-loop.

very interesting you mentioned human-in-the-loop "today" because it probably won’t be required in the future.

part of why karpathy’s autoresearch works well is because it is optimizing a single clean objective: validation bits per byte.

whereas an agent harness is much messier. you end up having to measure and tune multiple things at once like trajectory correctness, tool call success rate, end outcome quality, handoff efficiency, context groundedness, and a bunch of other interacting metrics.

[Hamel Husain @HamelHusain]

Made this video to explain evals

[Jainit Purohit @mjainit]

agents are commodity

edge = context quality × harness design × eval loops

compounding comes from faster:
trace → failure → tweak → re-run cycles

[Jainit Purohit @mjainit]

@Vtrivedy10 are you guys thinking about trace —> eval —> harness improvement loops in langsmith? feels like missing infra for fast harness iteration

[Jainit Purohit @mjainit]

another solid piece by @vtrivedy10

been a fan since he coined haas

agent performance is now as much a harness problem as a model problem

same model, different harness, wildly different outcomes

we’re still early. a lot of the alpha is still here

[Viv @Vtrivedy10]

https://t.co/Lb2gtM9SNs

[Jainit Purohit @mjainit]

@Vtrivedy10 agreed. big unlock now is trace-driven iteration

not just measuring runs, but learning from traces what and how to tweak in harness to hill climb fast

building with deepagents around this. measure, eval, tweak, repeat for long-horizon data tasks

[Alex Mordvintsev @zzznah]

Every time you read or listen to something, you're running untrusted code on your wetware with no sandbox.

Reading is code execution. Text is the oldest exploit.

Choose your inputs carefully. Including this post.

[Viv @Vtrivedy10]

https://t.co/NhTTz2UbKt

[ThePrimeagen @ThePrimeagen]

Before ai, you were bad decisions you could feel how bad they were and often could correct course on the after a couple weeks. Slowing down often helped make decisions I am happy about for months or years

Now our brick laying meme is the norm