What We Offer:

• Canteen Subsidy
• Night Shift allowance as per process
• Health Insurance
• Tuition Reimbursement
• Work-Life Balance Initiatives
• Rewards & Recognition

Key Responsibilities

• Fine-tune Llama 3.1 8B using LoRA/QLoRA on curated MI conversation datasets — 200+ MI scripts, recovery dialogues, crisis protocols. Ensure bounded, MI-consistent outputs aligned with recovery-oriented therapeutic principles
• Fine-tune Whisper Large-v3 on SUD-domain audio datasets — medical terminology, substance names, recovery vocabulary, slang/colloquialisms common in SUD populations
• Train or fine-tune StyleTTS2/Piper TTS to produce a single, clinically approved voice persona — warm, empathetic, gender/age appropriate as determined by clinical advisory board
• Curate, clean, and structure fine-tuning datasets in collaboration with the Clinical Advisor — ensure data is de-identified, clinically validated, and demographically representative
• Design and run hyperparameter optimization experiments — learning rate schedules, LoRA rank selection, training epoch optimization. Rigorous experiment tracking via MLflow
• Apply model compression for production — GPTQ/AWQ quantization for LLM, INT8 for ASR. Benchmark quality loss vs. latency/VRAM gains
• Build automated evaluation pipelines — LLM: MI consistency scoring, safety testing, hallucination rate. ASR: WER on domain test sets. TTS: MOS estimation, pronunciation accuracy
• Implement A/B testing infrastructure for model version comparison in pre-production
• Create comprehensive model cards documenting training methodology, data composition, performance, limitations, and bias audit results
• Collaborate with ML Engineers for seamless handoff of fine-tuned models to production inference pipelines

Required Qualifications

• 4+ years ML engineering with deep expertise in fine-tuning large language models
• Hands-on experience with LoRA, QLoRA, or full parameter fine-tuning of open-source LLMs (Llama, Mistral, Phi)
• Experience with model quantization techniques (GPTQ, AWQ, GGUF, bitsandbytes) and quality/performance trade-offs
• Strong proficiency in PyTorch, HuggingFace Transformers, PEFT, and training infra (DeepSpeed, FSDP)
• Experience fine-tuning speech models (Whisper, wav2vec2) on domain-specific audio data
• Strong understanding of training data curation — cleaning, deduplication, quality filtering
• Experience with systematic experiment tracking (MLflow, Weights & Biases)
• Solid evaluation methodology — designing test sets, avoiding data leakage, statistical significance
• NVIDIA GPU expertise — multi-GPU training, mixed precision (FP16/BF16), gradient checkpointing

Required Qualifications:

• Experience fine-tuning or training TTS models (StyleTTS2, VITS, Tacotron, or Piper)
• Experience with RLHF, DPO, or other alignment techniques for LLMs
• Prior work in healthcare or clinical NLP domains
• Experience building custom evaluation metrics for domain-specific NLP tasks
• Familiarity with Motivational Interviewing or therapeutic conversation patterns
• Experience with synthetic data generation for training data augmentation
• Understanding of bias auditing in ML models across demographic groups