Eekee AI is building the first proactive, pre-EAP mental-health layer for the modern workplace a private AI that helps employees build meaning, clarity, and resilience before burnout becomes a crisis.
We're not building another meditation app.
We're not building another reactive support tool.
We're building the product that sits before all of that the first line of mental-health engagement inside companies.
Role Description
Own training, alignment, and deployment of our meaning-at-work coach. You'll design data pipelines, fine-tune models, build safety layers, set evals, and ship to production.
What You'll Do
- Build data flow: sourcing, cleaning, consent/PII scrubbing, augmentation, labeling guidelines
- Fine-tune/align (SFT, DPO/RLHF) for coach behavior: listen reflect one powerful question next step
- Implement safety: crisis/HR escalation classifiers, toxicity/harassment filters, privacy guardrails
- Add context: retrieval for work scenarios; prompts/tools for Question of the Day and Vent
- Create evals: conversation quality, protocol adherence, safety, bias, latency, cost
- Optimize inference: distillation, quantization, caching, batching; observable, canaried rollouts
- Document: model cards, red-team results, alignment notes
Minimum Qualifications
- 1+ years applied ML/LLMs; shipped fine-tuned models
- Strong in PyTorch/JAX and serving (vLLM/Triton/Ray/K8s)
- Experience with SFT/DPO/RLHF, synthetic data, eval harnesses
- Built safety/quality classifiers and RAG systems
- Pragmatic on latency/cost; solid profiling chops
- Clear technical writing
Nice to Have
- Coaching/org-psych or agent-design background
- On-device/edge inference; multilingual safety
- Model cards, red-teaming, compliance awareness
- Zero-to-one startup experience
Stack
Models: Gemma/Anthropic, Llama/Mistral, LoRA/QLoRA, GGUF
Tooling: PyTorch, HF, vLLM, Triton, Ray, W&B
RAG: embeddings + FAISS/pgvector, guardrails
Infra: K8s, cloud GPUs, Terraform; analytics + experiment tracking
30/60/90 Day goals
- 30: Offline eval harness; baseline fine-tune matching coach style; first safety classifiers; model card v0
- 60: Prod daily question + vent flows w/ guardrails; cut token cost by 3050%; A/B live
- 90: Distilled model in prod; red-team playbook; documented alignment/escalation; +20% conv-quality score
