Joined March 2026
- Tweets 148
- Following 100
- Followers 60
- Likes 54
2 Photos and videos
Thought LLMs were just prompt→response
Then I dug deeper
• Why are production systems so much faster?
• Why memory becomes a bottleneck?
→ Continuous batching
→ Paged attention
→ Speculative decoding
Harder problems begin underneath:
latency, memory, throughput, serving.
1
2
12
Jun 11
Thought research reproduction meant:
Read paper → run code → verify results.
Reality 🥲
Working through BioReason (DNA Qwen GRPO):
broken imports, version mismatches, flash-attention/NVCC issues, Slurm debugging, config tuning…
Fix one thing → another breaks.
1
2
61
Jun 2
Coding with AI still feels broken.
Repo in mind.
Error visible.
Yet we stop…
to type explaining context again.
Why still code like this?
Been building Voker: a repo-aware voice coding agent.
AI going in wrong direction?
Just speak again.
No restart.
No prompt.
No broken flow
3
5
117
May 17
Went to verify a paper's results, I was like
"Yeah! Let's do something fun today🫠"
Reality:
- random dependency issues
- glibc confusion
- disk quota errors
Now, my hair is messy, eyes are sleepy and here I'm questioning my life choices 😭.
2
118
Apr 21
RAG can retrieve “relevant” chunks
and still miss the exact answer.
Similarity ≠ correct context.
Compared vector RAG vs PageIndex
(cached indexing → retrieval only)
Different retrieval → different answers.
Try it:
vidhan66-compare-rag.hf.spac…
3
117
Apr 12
Everyone’s talking about OpenClaw.
I tried building a 2-agent setup with a validator skill.
Skill was:
- registered
- visible
But never applied.
In LLM frameworks:
registration ≠ execution
Full version: linkedin.com/posts/vidhan-ba…
4
84
Apr 4
Most people use routing in LLM systems for intent:
→ query type → pipeline
But the real use is when the system isn’t confident:
→ retry
→ ask follow-ups
→ expand retrieval
Routing = handling failure states, not just intent.
2
47
Mar 31
Most notes are useless.
You read → don’t get it → go to YouTube.
I’m building:
Upload notes/PDF → AI generates a short whiteboard teaching video.
Trying to validate if this is actually useful.
1 min form:
forms.gle/NhRa1CELwKTUG4k57
Brutal honesty > fake validation.
1
32
Mar 28
Most LLM apps fail here:
They answer incomplete queries.
Better approach:
Detect missing fields
→ ask targeted questions
→ then answer
Not the other way around.
1
24
Mar 21
Why refusal logic is necessary in LLM systems —
One thing I noticed while building LLM systems:
The model almost always answers even when it shouldn’t. That’s the core issue behind hallucinations.
1
1
28
Mar 21
For scaling:
- ANN-based indexing becomes important as the dataset grows (I haven’t implemented yet)
Takeaway:
Hallucination control isn’t just a model problem. It’s a system design problem. If you don’t define when to refuse, your system will always answer even when it’s wrong.
1
1
21
Mar 21
Next, I’ll share how I handle ambiguity and make LLMs ask better follow up questions instead of answering too early.
If you are also into this field, I would love to know your approach and where I can improve it.
1
19
Mar 14
How my first LLM app turned into a maintenance nightmare —
At the start everything lived in one file:
• prompts
• retrieval logic
• API calls
• tools
• database queries
It worked… until the system started growing.
1
1
16
Mar 14
But prompts were still embedded in code.
That made iteration painful.
Every prompt change required touching application logic.
The final improvement:
Externalizing prompts into YAML configs.
Now prompt changes didn't require modifying the core system.
1
1
16
Mar 14
Big realization:
LLM apps behave more like systems engineering problems than simple scripts.
Modularity becomes important much earlier than expected.
1
16