Software Engineer · GenAI & ML · E2open

Kishor Hange

I build production LLM systems (RAG pipelines, AI agents and fine-tuned models) and make them fast, measurable and reliable once real users depend on them.

maximize p(useful | production)

See the work Résumé

Shipping ML and LLM systems at E2open: 3+ yrs
Lower inference latency via quantization & pruning: 38%
Accuracy over a legacy text classifier: +9.5%
Exact tool calls from a fine-tuned agent, up from 16.7%: 94.8%

01Selected work

Models I trained, measured and shipped in public

Alongside production work at E2open, I train and evaluate models end to end on cloud GPUs. Every number below links to code and raw results on GitHub.

Base LLM → SFT → DPO → benchmarks

Mini-ChatGPT

View on GitHub

The ChatGPT recipe end to end at a scale that fits one GPU: take a base model that can only continue text, teach it to be an assistant, align it to human preferences, and measure every stage against the model builder's own chat model.

Supervised fine-tuning of Qwen2.5-0.5B on 160k conversations (127M tokens) with assistant-only loss masking and length bucketing.
Direct Preference Optimization written from scratch (no TRL) on 12k UltraFeedback pairs: held-out preference accuracy 52% → 68%.
Evaluated with EleutherAI lm-evaluation-harness, the tool behind the Open LLM Leaderboard.

PyTorch
Transformers
SFT
DPO
lm-eval-harness
A100

Benchmarks: our 0.5B assistant vs Qwen's official one

Same base model, scores in %. Hover or focus a row for all three values.

Base model (Qwen2.5-0.5B)Ours: SFT + DPOQwen2.5-0.5B-Instruct (official)

View as table

Benchmark	Base	Ours	Official
IFEval (instruction)	22.7	36.7	36.3
IFEval (prompt)	12.6	24.0	26.4
ARC-Easy	58.4	64.3	59.2
ARC-Challenge	32.3	33.8	33.2
HellaSwag	52.2	52.0	52.4
MMLU	47.5	45.9	45.8

8 LLM projects · each against a measured baseline

Generative AI Engineering

View on GitHub

Fine-tuning, reinforcement learning, retrieval, agents, multimodal models and LLM safety, each trained on cloud GPUs, scored with a metric that matches the task, and honest about what still fails.

Area	Project	Result	Note
Fine-tuning	QLoRA text-to-SQL (Qwen2.5-7B)	Execution accuracy 60.0% → 66.2%	0.53% of params trained
Fine-tuning	Distilling 7B into 0.5B	Student 43.4% → 52.2%	no human labels
RL	GRPO for math reasoning	Format compliance 55% → 98%	accuracy flat, analysed
Agents	Function-calling LLM	Exact tool calls 16.7% → 94.8%	100% valid JSON
RAG	Retrieval, measured	nDCG@10 0.403 vs 0.217 (BM25)	reranker hurt, measured
Multimodal	Whisper for Marathi	CER 53.6% → 14.7%	11.9 h of speech
Multimodal	Vision-language chart QA	76.2% → 77.5%	within noise, reported
Safety	Prompt-injection guardrail	F1 0.997 in-distribution	14% recall on unseen attacks

QLoRA
PEFT
TRL
GRPO
FAISS
Whisper
Qwen2-VL

14 projects · vision, NLP, generative models

Deep Learning Portfolio

View on GitHub

Architectures built and trained from scratch or fine-tuned, each a single heavily commented train.py with its real metrics, figures and failure cases.

4.8× fasterBERT via distillation + ONNX + int8, 85% smaller

−69% memoryGPT-2 LoRA vs full fine-tuning

31 msAdaIN style transfer for any unseen style

0.60 mIoUU-Net semantic segmentation

77.9%ShiftViT: a ViT without attention, CIFAR-10

73% validMolecule generation with a SMILES GPT

PyTorch
Keras 3
ONNX Runtime
ViT
LoRA

Final-year project~12× smaller

CNN Edge Optimization

Compiled a Keras CNN for edge devices through TensorFlow Lite, ONNX and Apache TVM pipelines, checking accuracy at every step.

TFLite
ONNX
Apache TVM

View on GitHub

Streamlit app8 tokenizers

LLM Token Visualizer

Shows how BERT, GPT-2, GPT-4, Flan-T5, StarCoder2, Phi-3, Qwen2-VL and DeepSeek-R1 tokenizers split the same text, with token counts.

Hugging Face
Tokenizers
Streamlit

View on GitHub

Character-level RNN18 languages

Surname Origin Classifier

Predicts a surname's language of origin letter by letter; exported with TorchScript and served through a Streamlit app.

PyTorch
RNN
TorchScript

View on GitHub

02Approach

How I build ML systems

The same rules in production and in my own projects: a result only counts if it beats a baseline on the metric that matters.

Baseline first

Every fine-tuned model is compared with the same model zero-shot, or a classical method, on the same held-out data and prompt.

Metrics that match the task

SQL is executed and result sets compared; retrieval uses nDCG and recall; speech uses WER/CER; security uses out-of-distribution recall.

Show the failures

A reranker that hurt, a guardrail that misses unseen attacks, an MMLU drop after SFT: reported, not hidden.

Cost-aware by default

Hard time budgets per run, LoRA/QLoRA so one GPU is enough, and resumable checkpoints when cloud GPUs disappear.

03Experience

Where I've shipped

E2open

Bengaluru · Jul 2023 – Present

Software Engineer
Sep 2025 – Present
Associate Software Engineer
Jul 2023 – Sep 2025

Generative AI & LLMs

Built an LLM-powered chatbot that automatically resolves customer support tickets, reducing manual support workload.
Built RAG pipelines with LangChain, FAISS and Chroma, evaluated with Ragas and DeepEval for faithfulness, relevance and context precision.
Fine-tuned LLMs with PEFT (LoRA, QLoRA) for domain-specific and multilingual use cases.
Designed agentic workflows with LangChain, CrewAI and n8n, and implemented Model Context Protocol (MCP) for secure LLM access to enterprise databases.
Built semantic document matching on Vertex AI with Universal Sentence Encoder and Sentence Transformers.

ML systems, optimization & MLOps

Cut model inference latency by 38% with quantization and pruning (ONNX Runtime, Hugging Face Optimum), saving one full sprint of delivery time.
Fine-tuned BERT for business-document classification, and replaced a legacy classifier with a Random Forest on custom Word2Vec embeddings, improving accuracy by 9.5%.
Built real-time model monitoring and a feedback loop with FastAPI, Airflow, Prometheus and Grafana.
Built Spark, Hadoop and Kafka data pipelines with Snowflake analytics, and automated onboarding and operations workflows.

D-Smart Innovative Technology

Internship · Nov 2021 – Feb 2022

Web Development Intern
Nov 2021 – Feb 2022

Built reusable UI components with React.js, Bootstrap, jQuery and Sass; cut page load time by ~5% by removing unused CSS.

04Toolkit

What I work with

GenAI & LLMs

LangChain
LangGraph
CrewAI
MCP
RAG
FAISS
Chroma
Ragas
DeepEval
OpenAI API
AWS Bedrock

LLM training

Transformers
PEFT
LoRA / QLoRA
TRL
SFT
DPO
GRPO
Distillation
lm-eval-harness

ML & deep learning

PyTorch
TensorFlow
Keras
scikit-learn
XGBoost
BERT
ViT
Whisper

Optimization

Quantization
Pruning
ONNX Runtime
HF Optimum
TFLite
Apache TVM

MLOps

FastAPI
Docker
Kubernetes
GitHub Actions
MLflow
Comet ML
Airflow
Prometheus
Grafana

Data & cloud

Python
SQL
PySpark
Kafka
Snowflake
GCP Vertex AI
AWS SageMaker

Education

B.Tech, Computer Science & Engineering

DKTE's Textile & Engineering Institute, Ichalkaranji

2019 – 2023 · CGPA 8.31

Certifications

DeepLearning.AI (Coursera): Machine Learning, NLP, CNNs and Time Series in TensorFlow
AWS Academy Cloud Foundations

05Contact

Let's build something useful.

Open to GenAI, LLM and ML engineering roles, and always happy to talk about evaluation, RAG or fine-tuning. Email is the best way to reach me.

kishorhange46@gmail.com

Download résumé LinkedIn GitHub