Headline

Student at Rajiv Gandhi Institute of Petroleum Technology (RGIPT)| Asynq | CodingJr

About

I’m a Computer Science engineer focused on LLM inference, model optimization, and AI systems. My recent work includes: • Implementing Llama-3.2-1B from scratch in PyTorch, including GQA, RoPE, KV cache, SwiGLU and inference optimizations. • Building GPTQ and AWQ quantization from scratch, including packed INT4 weights and activation-aware scaling. • Analyzing inference performance using roofline modeling across T4, L4 and A100 GPUs. • Worked on LLM fine-tuning, knowledge distillation, and ML systems during internships at Siemens, Asynq Technologies, and Coding Jr. I’m particularly interested in LLM inference, quantization, GPU optimization, inference serving, and ML systems. GitHub: github.com/ankur110

Location

Patna

AI stack

claude code

deepmind

hugging face

ollama

llama.cpp

vLLM

Work experience

Machine Learning Researcher at Siemens, Mar 2026 - Jul 2026 · 5 mos, Implemented and benchmarked 7 hybrid physics + deep-learning architectures (3 Series/Parallel, Caspari 4-step, PINN, PINNverse, plus 1 novel) in PyTorch across 2 industrial case studies from 4 papers. Trained 6 crystallization model variants — 3 architectures × 2 backbones (681K-param Transformer, 282K-param LSTM) — on 520 simulated runs, reaching R² > 0.99 on batch dynamics. Built a 4-step pipeline (CasADi/IPOPT → feature selection → 8 MLPs → LSODA rollout) for a real pharmaceutical reactor (3 states, 2 inputs, 8 unknown parameters), hitting R² 0.9998 / 0.928 / 0.9999 on 2 unseen batches. Traced a non-identifiability bug inverting feed→temperature response; 1 physics-motivated input restriction lifted test R² 0.859 → 0.924 with zero retraining elsewhere. Designed a novel end-to-end differentiable architecture merging 2 stages into 1 loop (BPTT through differentiable RK4, physics-restricted inputs); filed as a Siemens invention disclosure. Recovered 5 kinetic parameters from 25 measurements at 10% noise via 8,192 Sobol collocation points; MDMM cut recovery error 13% → 4% and runtime 25 → 20 min. Reproduced a 6-agent LLM system (AutoGen, GPT-5) auto-generating MPC controllers on the BOPTEST benchmark — 340 passing tests, multi- vs single-agent success rate compared across repeated runs.

LLM Engineer Intern at Asynq, Jun 2025 - Aug 2025 · 3 mos, Remote, Architected ML pipelines for 3 client projects, dedicating 60%+ of resources to the flagship Sales Agent system. Fine-tuned Transformer models for spam classification, achieving 95% precision via targeted hyperparameter optimization. Engineered a deterministic fallback algorithm to handle API failures, ensuring 99.9% decision logic reliability. Implemented cosine similarity search for vector retrieval, enabling real-time updates to embeddings from user feedback. Designed human-in-the-loop validation loops to capture rejection data, automatically enriching the training dataset. Optimized inference orchestration using LangGraph, reducing latency by 40% for the Employee Assessment tool.

Machine Learning Engineer at Coding Jr, Jun 2024 - Aug 2024 · 3 mos, Fine-tuned 10+ LLMs (Mistral, Llama 2) via QLoRA and PEFT, and curated 10k+ de-duplicated instruction samples for robust supervised fine-tuning (SFT). Stabilized multi-GPU training with BF16 mixed precision and gradient accumulation on consumer GPUs; benchmarked SOTA architectures on perplexity and syntax accuracy.

Education

Bachelor of Technology - BTech, Cse, Rajiv Gandhi Institute of Petroleum Technology (RGIPT)

Public links

github: https://github.com/ankur110