Why this role
We're building agentic AI for recruitment workflowssourcing, screening, interview assistance, and offer orchestration. You'll own LLM/agent design, retrieval, evaluation, safety, and targeted traditional ML models where they outperform or complement LLMs.
What you'll do
- Hands-on AI (7080%): design & build agent workflows (tool use, planning/looping, memory, self-critique) using multi-agent frameworks (e.g., LangChain, LangGraph; plus experience with similar ecosystems like AutoGen/CrewAI is a plus).
- Retrieval & context (RAG): chunking, metadata, hybrid search, query rewriting, reranking, and context compression.
- Traditional ML: design and ship supervised/unsupervised models for ranking, matching, dedup, scoring, and risk/quality signals.
- Feature engineering, leakage control, CV strategy, imbalanced learning, and calibration.
- Model families: Logistic/Linear, Tree ensembles, kNN, SVMs, clustering, basic time-series.
- Evaluation & quality: offline/online evals (goldens, rubrics, A/B), statistical testing, human-in-the-loop; build small, high-signal datasets.
- Safety & governance: guardrails (policy/PII/toxicity), prompt hardening, hallucination containment; bias/fairness checks for ML.
- Cost/perf optimization: model selection/routing, token budgeting, latency tuning, caching, semantic telemetry.
- Light MLOps (in-collab): experiment tracking, model registry, reproducible training; coordinate batch/real-time inference hooks with platform team.
- Mentorship: guide 23 juniors on experiments, code quality, and research synthesis.
- Collaboration: pair with full-stack/infra teams for APIs/deploy; you won't own K8s/IaC.
What you've done (must-haves)
- 810 years in software/AI with recent deep focus on LLMs/agentic systems plus delivered traditional ML projects.
- Strong Python; solid stats/ML fundamentals (bias-variance, CV, A/B testing, power, drift).
- Built multi-agent or tool-using systems with LangChain and/or LangGraph (or equivalent), including function/tool calling and planner/executor patterns.
- Delivered RAG end-to-end with vector databases (pgvector/FAISS/Pinecone/Weaviate), hybrid retrieval, and cross-encoder re-ranking.
- Trained and evaluated production ML models using scikit-learn and tree ensembles (XGBoost/LightGBM/CatBoost); tuned via grid/Bayes/Optuna.
- Set up LLM and ML evals (RAGAS/DeepEval/OpenAI Evals or custom), with clear task metrics and online experiments.
- Implemented guardrails & safety and measurable quality gates for both LLM and ML features.
- Product sense: translate use-cases into tasks/metrics; ship iteratively with evidence.
Nice to have
- Re-ranking (bi-encoders/cross-encoders), ColBERT; semantic caching; vector DBs (pgvector/FAISS/Pinecone/Weaviate).
- Light model serving (vLLM/TGI) and adapters (LoRA); PyTorch experience for small finetunes.
- Workflow engines (Temporal/Prefect); basic time-series forecasting; causal inference/uplift modeling for experiments.
- HRTech exposure (ATS/CRM, interview orchestration, assessments).
