❯ ai engineer · otsuka corporation · iit guwahati '24 · tokyo 🇯🇵

Hello, I'm .

I build agentic systems, fine-tune LLMs, and get RAG working on real, messy data. Most of my time goes to the gap between a promising demo and something people can actually rely on.

finetune/run_qlora.py
1 2 3 4 5 6 7 8 9
from peft import LoraConfig from trl import SFTTrainer trainer = SFTTrainer( model=base_model, # 10+ LLMs tuned train_dataset=sft_pairs, # SFT · RLHF · LoRA peft_config=LoraConfig(r=16), ) trainer.train()
~/eval
$ 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 build and run production LLM systems — RAG chatbots on Elasticsearch, FAQ and quotation assistants with hybrid search and custom MCP tools, and fine-tuning pipelines across multi-node on-prem GPUs.

I studied at IIT Guwahati and work across the full ML lifecycle: data curation, training, RLHF, evaluation, and deployment. I like the unglamorous parts — reproducible baselines, honest evals, and pipelines that don't fall over. Living in Tokyo, JLPT N4 certified.

trained on H100A100A6000
❯ in production
10+
LLMs fine-tuned with SFT, RLHF & LoRA/QLoRA
85+
composite eval score on JA-centric models
+31%
retrieval accuracy on deployed search
−23%
search time after hybrid retrieval
❯ /experience

Experience.

Oct 2024 — PresentOtsuka Corporation · Tokyo
AI Engineer — Full Time
  • Lead work on domain-adapted LLMs on multi-node GPUs (H100 / A100 / A6000) — LoRA training, deployment, and RAG over the company's own data.
  • Run fine-tuning, SFT, instruction tuning, and RLHF (PPO, DPO, ORPO) with Axolotl and LLaMA-Factory, comparing LoRA/QLoRA vs. full fine-tuning to keep baselines honest.
  • Set up evaluation with DeepEval (MT-Bench, TruthfulQA, custom checks); JA-centric models clear 85+ composite scores.
  • Built the RAG chatbots teams actually use — custom MCP tools + Elasticsearch, connectors, ingestion pipelines, JP-EN analyzers.
Jun 2023 — Jul 2023Otsuka Corporation · Tokyo
AI Engineering Intern
  • Recreated Stanford's “Generative Agents: Interactive Simulacra of Human Behavior” — agents that react, talk, remember, and decide inside a small game world.
  • Wired it up with OpenAI LLMs and LangChain, then spent a good while on prompt engineering to make behaviour believable.
  • Took it further into an Assistive Sales Agent prototype that handles client conversations more like a person would.
2021 — 2024earlier · internships & research
AI Intern
HODM
Dec 2023 — Jan 2024 · Delhi
Applied machine learning to digital-marketing problems.
SDE Intern
SanchiConnect
Dec 2022 — Jan 2023 · Noida
Two-way recommender matching deep-tech startups with investors + a candidate-scanning system for cloud HR.
Research Intern
DIAT, DRDO
Dec 2021 · Pune
Crowd-analysis research with Dr. Sunita Dhavale — surveillance + live video-streaming feeding DL models. (LOR.)
Web Developer
IIT Guwahati
Aug 2021 — Sep 2021 · Remote
Under Prof. Sastri, built alumni portals for UoH, NIPER Guwahati & IIT Guwahati.
❯ /projects

Projects.

Open-source toolkits from my own work with language models.

❯ ls more/ → StockData_ETL, TextBoat, dev.connect  ·  all repositories on GitHub →

❯ /skills

Skills.

llms & agents

LangChainLangGraphLlamaIndexMCP Hugging FacevLLMAxolotlLLaMA-Factory LoRA / QLoRARLHFRAG

ml / deep learning

PyTorchTransformersscikit-learn NumPyPandasWeights & Biases

programming

PythonCC++BashSQL

cloud & mlops

Azure AI FoundryMLflowLangFuse DeepEvalApache Airflow

databases & search

ElasticsearchPostgreSQLMongoDB MySQLFAISSChromaDB

infrastructure

DockerGitLinuxCI/CD
❯ /contact

Let's build something.

I'm always happy to talk about AI engineering, research, or new opportunities — in English or Japanese. If you're building something interesting in ML or AI, drop me a line.