一About自己紹介
Retrieval scores over self-reported certainty. Pruned weights over bloated checkpoints. Shipped code over benchmark screenshots.
So I go one layer down: I write the scheduler instead of calling generate(), and the autograd instead of importing it.
BIT Mesra. Still in a staring contest with one code block. We have beef.
full precision
… contributions in the last year · @TryingtobeingNikhil ↗
二Experience経歴
Where my code met real users.
Machine Learning Engineer – AI Infrastructure · Remote
Software Engineer, ML
Live product · dev-path.site
Roadmaps and one focused task a day, for developers stuck in tutorial hell. Built and shipped solo.
Model compression collaboration
Pruned a U-Net for biomedical segmentation: 97.3% fewer parameters, IoU still above 0.95. Most of it was along for the ride.
三Work作品10 projects
Two systems, end to end. One serves LLMs, one trains agents.
LLM inference & serving
A vLLM-style inference engine, grown from a naive generate() loop in 11 phases.
Continuous batching, a paged KV cache and CPU swap. Under load it slows down instead of falling over.
Reinforcement learning · C++20
A reinforcement-learning library with no import torch, down to its own autograd.
Tensors, autograd, DQN and PPO in C++20, zero dependencies, and the diamond-graph gradient that most from-scratch engines get wrong.
PPO, threading and CUDA backends co-built with aprv10.
$ cmake --build build -j && ctest 100% tests passed, 0 failed out of 148 $ ./train --env gridworld --steps 20000 eval return -100.0 → 3.0 # optimal
A 5-agent pipeline (Planner → Researcher → Critic → Writer → Evaluator) that loops until a deterministic confidence threshold is met, computed from mean retrieval score, not the LLM's self-report. The same rule "Ask this page" uses up there.
Retrieval by LLM-guided tree traversal instead of embeddings. I wanted to know whether the vector-DB default was actually necessary, so I built the version without one.
Fine-tuned a 767M summariser with 1.57M trainable params. Full fine-tuning fell apart on unseen domains; LoRA didn't.
Research collaboration with IIT Kharagpur. Structured pruning of U-Net for biomedical segmentation: 97.3% fewer parameters, 92% fewer FLOPs, IoU still above 0.95 on MoNuSeg.
The full Transformer from the paper, multi-head attention to training loop. Not "used a framework."
Skin-lesion classification across 7 classes with a 57:1 imbalance. Focal loss so the rare classes count, Grad-CAM so you can see what it looked at, ONNX export so it runs anywhere.
A senior engineer code-reviews your brain. Six-stage multimodal pipeline, real models, no affirmations.
四Lab実験室5 toys
A small model drafts 4 tokens; the big one checks them in one pass and keeps what matches.
The KV cache stores the ____
Low sharpens, high flattens. Drag, then sample.
Too small crawls. Past 0.167 it diverges.
hover a token
Each word only looks back. Hover one.
The split GPT-2 makes before BPE, live. Your name costs more tokens than "the".
0 pre-tokens0 characters0 characters per token
五Writing文章20 articles
What I cannot create, I do not understand.
六Contact便り
Email is fastest. DMs on X work too.
It's --:-- in Jaipur (IST)