Looking for a job

Shukraditya Bose

@shukradityabose

Message

Machine Learning Engineer at Omli Kids building speech models and reducing inference costs by 4x.

India

Message

At Omli Kids Pvt. Ltd., I fine-tuned Qwen3-8B for children’s speech understanding and reduced serving costs by 40%. I also benchmarked inference engines and increased single-A100 concurrency from one to four requests while keeping latency targets in range.

At TIFIN India, I architected multi-agent workflows and implemented multimodal RAG pipelines, improving task success by 28% and reducing rule-based error rates by about 30%. I built Python-based LLM-as-a-judge frameworks that cut manual review effort by about 75%.

My work at Memfold AI, Sarvam AI, and Samsung R&D Institute included RAG systems, multilingual OCR for Indic scripts, and continual-learning experiments. In my projects, I’ve built a self-improving conversational agent and a Bengali subword tokenizer.

Experience

Work history, roles, and key accomplishments

Current

Machine Learning Engineer

Current

Omli Kids Pvt. Ltd.

May 2026 - Present (5 months)

Fine-tuned Qwen3-8B with LoRA for children's speech understanding and implemented AWQ BF16-weight compatibility fix to reduce serving costs by 40%. Benchmarked inference on vLLM, SGLang, and TensorRT-LLM, increasing single-A100 concurrency from 1 to 4 requests while keeping latency low.

Qwen3 8B LoRAs AWQ VLLM Sglang NVIDIA TensorRT LLM Voice Activity Detection Autoencoder LSTM ONNX Runtime

AI Engineering Intern

TIFIN India

Jan 2026 - May 2026 (4 months)

Architected multi-agent AI workflows with orchestration and handoff logic, improving task success by 28%. Implemented multimodal RAG pipelines and LLM-as-a-judge frameworks, reducing manual review effort by about 75%.

Multi Agent Systems RAG Multimodal AI LLM As A Judge Python

AI Engineering Intern

Memfold AI

Aug 2025 - Dec 2025 (4 months)

Developed a standalone RAG pipeline with Dockerized local vector database, integrating BM25 and dense embeddings with hybrid search and reranking. Built Slack, Discord, and Gmail bots using RAG services and analyzed interaction logs to improve retrieval behavior.

RAG BM25 Dense Embeddings Hybrid Search Reranking Docker Tiptap Slack API Discord APIs Google Mail API

Vision Language Model Intern

Sarvam AI

May 2025 - Aug 2025 (3 months)

Architected end-to-end multilingual OCR pipeline for Indic scripts, orchestrating data transformation across Qwen 2.5VL and Gemma 3 models. Engineered Policy Optimisation reward systems and fine-tuned YOLO models for Gujarati handwritten text recognition, improving accuracy by 15%.

OCR Qwen2.5 Gemma 3 YOLO PyTorch Hugging Face Langfuse Clickhouse Knowledge Graphs RAG

Virtual Intern

Samsung R&D Institute

Jan 2025 - Jun 2025 (5 months)

Developed an adaptive regularization algorithm for continual learning using Fisher information matrices. Integrated online EMA updates to improve resistance to catastrophic forgetting and benchmarked against EWC and MAS baselines.

Continual Learning Fisher Information EMA PyTorch EWC MAS

Education

Degrees, certifications, and relevant coursework

Vellore Institute of Technology

Integrated M.Tech., Computer Science with Specialisation in Data Science

Grade: 8.87/10.00

Activities and societies: Department Merit List

Pursuing an Integrated M.Tech. in Computer Science with Specialisation in Data Science, with a cumulative GPA of 8.87/10.00 and recognition on the Department Merit List.