❯chakshu · ai engineer · tokyo

Hello, I'm Chakshu.

I'm an AI engineer at Otsuka Corporation in Tokyo. I build RAG pipelines, multi-agent assistants, and fine-tuned LLMs that people use every day.

Get in touch View GitHub
finetune/run_qlora.py 
1from peft import LoraConfig
2from trl import SFTTrainer
3
4trainer = SFTTrainer(
5    model=base_model,        # 10+ LLMs tuned
6    train_dataset=sft_pairs, # SFT · RLHF · LoRA
7    peft_config=LoraConfig(r=16),
8)
9trainer.train()
$ python eval/run_benchmarks.py
✓ retrieval accuracy   +31%
✓ search latency      −23%
✓ queries automated   60%
running in production · Tokyo, JP
$
❯/about

About me.

I'm an AI Engineer at Otsuka Corporation in Tokyo, where I design and run production LLM systems: RAG chatbots on the ELK stack, quotation and FAQ assistants with hybrid search and multi-agent orchestration, and fine-tuning pipelines across multi-node on-prem GPUs.

I studied Artificial Intelligence & Data Science at IIT Jodhpur, specializing in Computer Vision, and my work covers the full ML lifecycle: data curation, training, evaluation, and deployment.

I also enjoy teaching. I was a Teaching Assistant for machine learning courses at IIT Jodhpur, and helped organize TEDx there.

Living in Japan, I hold a Japanese language certification (NAT-TEST 5Q) and work day-to-day with teams in Tokyo.

❯/experience

Experience.

Oct 2024 – PresentTokyo, Japan

AI Engineer · Otsuka Corporation

  • Built and deployed RAG chatbots on the ELK stack, improving retrieval accuracy by 31% and cutting search time by 23%.
  • Scaled quotation and FAQ assistants with vector + hybrid search and multi-agent orchestration (memory + planning), automating 60% of routine queries.
  • Ran LLM fine-tuning (SFT, instruction tuning, RLHF) on 10+ models with LoRA/QLoRA across multi-node on-prem GPUs, validated against standardized benchmarks.
  • Working with knowledge graphs and vector databases to improve retrieval, recommendations, and search.
Jun – Sep 2024Remote

Data Scientist · Laika AI

  • Architected a multimodal LLM chatbot (text + on-demand charts) with custom blockchain tools using AutoGen, improving task completion by 13% on internal benchmarks.
  • Refined data access and caching with Redis, cutting external API calls by 20% and chatbot response time by 50%.
Jun – Aug 2023Tokyo, Japan

AI Intern · Otsuka Corporation

  • Built an LLM agent simulation based on the Werewolves of Miller Hollow game, where multiple agents plan, act, and converse in real time.
  • Developed an application for group conversations with multiple AI agents, and an evaluation pipeline scoring agent plans and actions on success, safety, and latency.
Aug 2022 – May 2023IIT Jodhpur

Teaching Assistant · CSE, IIT Jodhpur

  • Teaching assistant for Introduction to Machine Learning and Pattern Recognition & Machine Learning, with teaching and assessment duties for 250+ students in each course.
Jun – Jul 2022Remote

Machine Learning Intern · Deep Learn Strategies

  • Implemented ML models for 3D reconstruction, Temporal Fusion Transformers, and face spoofing detection.
❯/projects

Projects.

Personal and research projects spanning LLM data pipelines, medical imaging, and 3D vision.

llmsft · rlhfdata pipeline

SFT & Alignment Dataset Builder

A toolkit that turns Markdown notes into structured SFT/RLHF datasets, with LLM-guided chunking, multi-turn QA synthesis, ORPO-style preference pairs, and quality gates with deduplication. Removes the manual labeling bottleneck for domain fine-tuning.

view repository →
computer visionresearchmedical imaging

Retinal Disease Detection

Led a 3-member research team at IIT Jodhpur on a retinal screening pipeline covering vessel segmentation, enhancement, and multi-disease classification. Benchmarked CNNs against Vision Transformers and added Grad-CAM interpretability for clinician trust.

view repository →
3d visionnerfresearch

3D Reconstruction

Implemented Neural Radiance Fields to generate 3D objects from multiple 2D views and camera poses, and used PIFuHD to reconstruct 3D human models from single images, putting modern reconstruction methods into practice.

view repository →

all repositories on GitHub →

❯/skills

Skills.

llms & agents

LangChainLlamaIndexMCPAutoGenCrewAIHugging FacevLLMLoRA / QLoRARLHFRAG

ml / deep learning

PyTorchTensorFlowscikit-learnOpenCVNumPyPandasWeights & Biases

programming

PythonCC++BashSQL

cloud & mlops

AWSBedrockSageMakerGCPVertex AIAzure AI FoundryMLflowLangFuseDeepEval

databases & search

ElasticsearchPostgreSQLMongoDBRedisNeo4jFAISSChromaDB

infrastructure

DockerKubernetesGitLinuxCI/CDNginx

Chess

FIDE rated 1551. Captained the chess team at IIT Jodhpur.

Teaching

Teaching assistant for machine learning courses (IML, PRML) at IIT Jodhpur. I enjoy helping others learn.

TEDx

Helped organize and manage TEDx IIT Jodhpur.

Japanese

NAT-TEST 5Q certified in grammar, reading, and listening.

❯/contact

Get in touch.

I'm always happy to talk about AI engineering, research, teaching, or new opportunities, in English or Japanese. Feel free to reach out.

Say hello