Vansh Kumar

Give me a difficult AI systems problem; I'll figure it out

[Vansh Kumar @k11107_kumar]

AI/ML Developer | AI Systems Builder
Building LLMs, RAG pipelines & AI Agents.
Exploring AI Research, ML & Agentic Systems.
Open to collaborations, research & projects.
Building in public. 🚀

[Vansh Kumar @k11107_kumar]

What if AI could design its own multi-agent architecture?
AgentFlow explores exactly this.
It optimizes agent roles,communication graphs tool access,message schemas and retry logic using runtime feedback.
The interesting part:optimizing the harness,not just the model.
#AI #Agents

[Vansh Kumar @k11107_kumar]

@Surendar__05 solid grinding going on bro 😎😭

[Vansh Kumar @k11107_kumar]

100% GPU usage doesn't mean your GPU is dying. 💀

My RTX 4050 hit repeated usage spikes, so I checked:

→ Temp: 66°C
→ VRAM: 3.3/6 GB
→ Current usage: 0%
GPU utilization ≠ GPU temperature ≠ VRAM usage.
Monitor thermals, not just Task Manager graphs.

#GPU #MachineLearning

[Vansh Kumar @k11107_kumar]

Why treat Claude Code like a chatbot when you can run it like a team?

1./init for repo context
2.Plan before edits
3.Delegate search to Haiku
4.Use stronger models for architecture
5.Test every change
6./compact when context grows

Orchestration > prompting.

#AI #ClaudeCode

[Vansh Kumar @k11107_kumar]

1-question I'm?
How should we eval an AI agent when the final answer is correct,but the execution path was inefficient?
Task success alone isn't enough.
We need to measure tool selection,recovery,latency, token cost and execution quality.
What's your approach to agent evaluation?

[Vansh Kumar @k11107_kumar]

AI is moving beyond chat.

But tool access ≠ autonomy.

Reliability means choosing tools, managing context, recovering from failure, and verifying outputs.

Making an agent act is easy. Making it reliable is hard.

#AI #AIagents

[Vansh Kumar @k11107_kumar]

Agentflow on work progress..🚀

[Vansh Kumar @k11107_kumar]

@i_mika_el True🥲

[Vansh Kumar @k11107_kumar]

Building an AI agent is easy.Building a reliable one is hard.
Giving an LLM access to tools doesn't automatically make it intelligent.
The real engineering challenges are:
→Planning
→Memory management
→Tool selection
→Error recovery
→Evaluation
Reliability > flashy demos. 🧠

[Vansh Kumar @k11107_kumar]

Hey @grok ,What's you opinion on this?

[Vansh Kumar @k11107_kumar]

Today's LeetCode POTD 🧩

Climbing Stairs II

Dynamic programming is more than just defining states—it's about identifying valid transitions under constraints.

Breaking a problem down, one state at a time. 🧠

#LeetCode #DSA

[Vansh Kumar @k11107_kumar]

@iamnagabandi @Raghulchan2004 Bro the material has given see that..
Notes were highly upto point.
No videos there but notes and quizzes were great..
Worth watching means the course material that type of material for reading and implementing directly from published paper can't you think of nice course?

[Vansh Kumar @k11107_kumar]

Today, many advanced coding agents depend on expensive infrastructure and large models. We want to investigate whether a smaller local model can become significantly more useful through intelligent orchestration.
👇

[Vansh Kumar @k11107_kumar]

Find Servers That Handled Most Number of Requests (Hard)
Key concepts: Priority Queue + Ordered Set + Greedy Simulation.
The interesting part isn't processing requests—it's efficiently finding the next available server while respecting cyclic assignment.
Optimize the search. 🧠

[Vansh Kumar @k11107_kumar]

For research work try this https://t.co/fKr1kgldWL

[Vansh Kumar @k11107_kumar]

@nandikagupta2 How does the system fuse the YOLO target position with the terrain map to decide move / turn / stop? Is it a simple rule-based state machine, potential fields, or something like DWA / pure-pursuit?

[Vansh Kumar @k11107_kumar]

just go through this article on 1 view
pov: your are ahead of everything after viewing this article.

[shikamaru @shikamarutsai]

https://t.co/M4Fa93bfTC

[Vansh Kumar @k11107_kumar]

Today's LeetCode POTD: 1267 — Count Servers that Communicate. 🧩

A good reminder that grid problems aren't always about traversal—sometimes row/column frequency counts are enough to identify servers that can communicate.

Think in counts before reaching for BFS.
#problemsolver