AI Engineer

As an AI Engineer specializing in Training & Alignment at eeKee AI in Bengaluru/Bangalore, India, you will be instrumental in the development and deployment of our meaning-at-work coach. This role involves end-to-end ownership of training, alignment, and production rollout. You will design robust data pipelines, meticulously fine-tune AI models, implement critical safety layers, establish comprehensive evaluation metrics, and ensure seamless delivery to production.

Key Responsibilities:

• Design and manage data flows, including sourcing, cleaning, PII scrubbing, augmentation, and developing labeling guidelines.
• Implement advanced fine-tuning and alignment techniques such as SFT, DPO, and RLHF to sculpt coach behavior, focusing on empathetic listening, reflection, providing impactful questions, and guiding next steps.
• Develop and integrate safety mechanisms, including crisis/HR escalation classifiers, toxicity/harassment filters, and privacy guardrails.
• Incorporate contextual information for work scenarios, and build prompts/tools for features like 'Question of the Day' and 'Vent'.
• Create and maintain evaluation frameworks to assess conversation quality, protocol adherence, safety, bias, latency, and cost.
• Optimize inference performance through techniques like distillation, quantization, caching, and batching, coupled with observable, canaried rollouts.
• Document key aspects of the models, including model cards, red-teaming findings, and alignment strategies.

Minimum Qualifications:

• At least 1 year of applied experience in Machine Learning and Large Language Models, with a proven track record of shipping fine-tuned models.
• Proficiency in PyTorch/JAX and model serving frameworks such as vLLM, Triton, Ray, or Kubernetes.
• Demonstrated experience with SFT, DPO/RLHF, synthetic data generation, and evaluation harnesses.
• Experience in building safety/quality classifiers and Retrieval-Augmented Generation (RAG) systems.
• A pragmatic approach to latency and cost optimization, with strong profiling skills.
• Excellent technical writing abilities.

Nice to Have:

• Background in coaching, organizational psychology, or agent design.
• Experience with on-device/edge inference or multilingual safety features.
• Familiarity with model cards, red-teaming, and compliance standards.
• Prior experience in zero-to-one startup environments.

Tech Stack:

• Models: Gemma/Anthropic, Llama/Mistral, LoRA/QLoRA, GGUF
• Tooling: PyTorch, Hugging Face, vLLM, Triton, Ray, Weights & Biases
• RAG: Embeddings + FAISS/pgvector, Guardrails
• Infrastructure: Kubernetes, Cloud GPUs, Terraform, Analytics & Experiment Tracking

30/60/90 Day Goals:

• 30 Days: Establish an offline evaluation harness, develop a baseline fine-tune aligning with coach style, implement initial safety classifiers, and produce a model card v0.
• 60 Days: Deploy 'Daily Question' and 'Vent' flows in production with guardrails, achieve a 30-50% reduction in token costs, and initiate A/B testing.
• 90 Days: Integrate a distilled model into production, create a red-teaming playbook, document alignment and escalation procedures, and achieve a 20% increase in conversation quality scores.