Blue Machines.ai - Senior Data Scientist - LLM Training & Fine-tuning

Job Title
Senior Data Scientist— LLM Training & Fine-tuning (Indian Languages, Tool Calling, Speed)
Location: Bangalore

About the Role

We’re looking for a hands-on Data Scientist / Research Scientist who can fine-tune and train open-source LLMs end-to-end—not just run LoRA scripts. You’ll own model improvement for Indian languages + code-switching (Hinglish, etc.), instruction following, and reliable tool/function calling, with a strong focus on latency, throughput, and production deployability.
This is a builder role: you’ll take models from research → experiments → evals → production.

What You’ll Do (Responsibilities)

• Train and fine-tune open LLMs (continued pretraining, SFT, preference optimization like DPO/IPO/ORPO, reward modeling if needed) for:
Indian languages + multilingual / code-switching
Strong instruction following
Reliable tool/function calling (structured JSON, function schemas, deterministic outputs)
• Build data pipelines for high-quality training corpora:
Instruction datasets, tool-call traces, multilingual data, synthetic data generation
De-duplication, contamination control, quality filtering, safety filtering
• Develop evaluation frameworks and dashboards:
Offline + online evals, regression testing
Tool-calling accuracy, format validity, multilingual benchmarks, latency/cost metrics
• Optimize models for speed and serving:
Quantization (AWQ/GPTQ/bnb), distillation, speculative decoding, KV-cache optimizations
Serve via vLLM/TGI/TensorRT-LLM/ONNX where appropriate
• Improve alignment and reliability:
Reduce hallucinations, improve refusal behavior, enforce structured outputs
Prompting + training strategies for robust compliance and guardrails
• Collaborate with engineering to ship:
Model packaging, CI for evals, A/B testing, monitoring drift and quality
• Contribute research:
Read papers, propose experiments, publish internal notes, and turn ideas into measurable gains

What We’re Looking For (Qualifications)

Must-Have

• 4 - 6 years in ML/DS, with direct LLM training/fine-tuning experience
• Demonstrated ability to run end-to-end model improvement:
data → training → eval → deployment constraints → iteration
• Strong practical knowledge of:
Transformers, tokenization, multilingual modeling
Fine-tuning methods: LoRA/QLoRA, full fine-tune, continued pretraining
Alignment: SFT, DPO/IPO/ORPO (and when to use what)
• Experience building or improving tool/function calling and structured output reliability
• Strong coding skills in Python, deep familiarity with PyTorch
• Comfortable with distributed training and GPU stacks:
DeepSpeed / FSDP, Accelerate, multi-GPU/multi-node workflows
• Solid ML fundamentals: optimization, regularization, scaling laws intuition, error analysis

Nice-to-Have

• Experience with Indian language NLP:
Indic scripts, transliteration, normalization, code-mixing, ASR/TTS text quirks
• Experience with pretraining from scratch or large-scale continued pretraining
• Practical knowledge of serving:
vLLM / TGI / TensorRT-LLM, quantization + calibration, profiling
• Experience with data governance: privacy, PII redaction, dataset documentation
Tech Stack (Typical)

• PyTorch, Hugging Face Transformers/Datasets, Accelerate
• DeepSpeed / FSDP, PEFT (LoRA/QLoRA)
• Weights & Biases / MLflow
• vLLM / TGI / TensorRT-LLM
• Ray / Airflow / Spark (optional), Docker/Kubernetes
• Vector DB / RAG stack familiarity is a plus

What Success Looks Like (90–180 Days)

• Ship a fine-tuned open model that measurably improves:
Instruction following and tool calling correctness
Indic language performance + code-switching robustness
Lower latency / higher throughput at equal quality
• Stand up a repeatable pipeline:
dataset versioning, training recipes, eval harness, regression gates
• Build a roadmap for next upgrades (distillation, preference tuning, multilingual expansion)

Interview Process

• 30-min intro + role fit
• Technical deep dive: prior LLM work (training/evals/production constraints)
• Take-home or live exercise: design an LLM fine-tuning + eval plan for tool calling + Indic language
• Systems round: training/serving tradeoffs, cost/latency, failure modes
• Culture + collaboration round