
MTS @ParsewaveAI | LLMs • AI Evals • Post-Training
Bhilai
Joined June 2025
- Tweets4.5K
- Following165
- Followers2.2K
- Likes12.3K
A week ago, if you told me I'd be working with people from OpenClaw and Hugging Face, I'd have laughed.
Funny how much can change in a week.
3
20
596
I have come across this paper by Apple. Their problem statement was whether RL can train agents to perform Long Horizon multi step tasks in interactive environment instead of teaching them from static examples.
They have used AppWorld as a simulated environment that contains multiple apps and APIs and trained Qwen2.5-32B-Instruct model under 24 scenarios (72 tasks).
The model was not given any instructions on how to solve each task and was left to interact with the environment and learn from the environment rewards based on task outcomes. In short they were using an RL method LOOP that doesn't use a value network and keeps only one copy of the underlying LLM in memory.
They trained a 32B agent which reached 71.3% task completion on the normal AppWorld test setting. Also OpenAI o1 scored 61.9% on the same setting, so the 32B agent beat o1 by about 9 percentage points.
The paper reports this as a 15% relative improvement.
So what's actually interesting in this was the agent's behaviour? It became more likely to read API documentation, made fewer unnecessary assumptions, reduced hallucination, and got better at recovering from setbacks and errors.
If I add a different perspective, the agent learned better interaction strategies instead of better answers.
The important finding is that they trained a relatively small agent in an environment, and RL changed the way the agent went about solving problems.
Paper: Reinforcement Learning for Long-Horizon Interactive LLM Agents by Kevin Chen et al.
1
8
360
My GF's first day at work 😊
She lasted approximately 9 hours before broadcasting the chaos at her office.
7
41
4,018
Today I solved Valid Sudoku from NeetCode 250.
I started with a complicated list slicing approach, then realized I needed to track rows, columns and boxes with hash sets. I initially created 27 sets manually cuz I'm afraid of defaultdict(set). But I finally learnt it.
10
503
Time to find out how updated my professors are with the current industry.
Mock interview season begins. Let's see what questions they come up with.
4 August 2026
Today's NeetCode 250 problem was Range Sum Query 2D - Immutable.
I started with the naive approach, but it quickly became obvious why it doesn't scale when there are lots of queries. Then I learned the 2D Prefix Sum technique, and it completely changed how I looked at the problem. Precompute once, answer every query in constant time.
On the backend side, I started learning Pytest. I practiced on a few sample files, then configured a separate test database for my FastAPI and began writing tests for my API functions.
12
434
Today's NeetCode 250 problem was Encode and Decode Strings.
I started with a simple delimiter based approach. Then I switched to length + delimiter encoding strategy. Such a simple problem, yet taught me a different perspective.
2
21
432
Today's NeetCode 250 problem was Top K Frequent Elements.
I solved it using a HashMap + sorting approach instead of Bucket Sort. Bucket Sort felt like an overkill for the problem but surprisingly it's the more optimal solution.
1
18
562
The beauty of persistence is that it quietly compounds.
I spent nearly 14 hours building this web scraper. I wanted to give up more times than I can count, but I kept debugging.
It's finally working.
I spent an entire day building a web scraper for a take home assessment.
Finally got everything working, and when I wanted to test it
The server responded with a 503 💀
3
19
1,104
My girlfriend is tired of doomscrolling and isn't finding motivation.
So I made her a small test. Handing it over in 2 hours.
4
21
818
30 July 2026
Today's NeetCode 250 problem was Sort Colors. My brute force solution was... Heap Sort
Then I looked at the optimal solution and discovered the Dutch National Flag algorithm. I thought, there's no way this gets any better. Then I came across another elegant in place approach. A single problem ended up teaching me multiple ways of thinking.
I didn't get to do backend today because I wanted to make my first open source contribution as part of a job application. I spent a few hours understanding the contribution workflow, found a project, and started exploring it.
Unfortunately, the project wasn't in a state I was comfortable contributing to. The codebase was largely AI generated, the website itself had several issues, and after spending the entire day trying to make sense of it, I decided to move on and look for a better project.
10
303
29 July 2026:
Solved Sort an Array from the NeetCode 250. I was overconfident and thought I'd figure out an efficient sorting algorithm from scratch. Reality humbled me pretty quickly. I went back, studied Heap Sort again, and after a lot of trial and error, finally got it working.
On the backend side, I learned how to manage environment variables, run Alembic migrations against a production database, serve as FastAPI application with Gunicorn, and create a systemd service so the application runs as a managed Linux service.
1
14
339
28 July 2026
Solved Design HashSet and Design HashMap from the NeetCode 250. There was a lot of brainstorming today. Implementing HashSet finally helped me understand what hashing is doing under the hood, and HashMap felt much more intuitive after that. I also wrote detailed explanations for both solutions in the discussion section.
On the backend side, I started working with Ubuntu VMs. Learned how to provision a VM and deploy my FastAPI application on it. Having previously worked on remote CentOS machines during my research made the whole process much easier.
17
391






















