AI Engineer

We're hiring an AI/ML engineer to own model intelligence end to end: making the core AI measurably smarter every week, and building the learning loops that turn production usage into continuous improvement.

What you will own

1. Applied LLM engineering. Prompt architecture, structured outputs, tool/function calling, context management, multi-step agentic flows — and the eval harnesses that prove each change is actually an improvement, not a vibe.
2. Continuous learning loops. Design the data flywheel: capture production interactions, label/score them (automated + human), and feed them back via fine-tuning, preference optimization (DPO/RLHF-style methods), or retrieval improvements — so the system gets better with every real-world use.
3. Retrieval and memory. Embedding pipelines, chunking strategy, hybrid retrieval, reranking, and long-horizon memory — measured by answer quality, not by architecture diagrams.
4. Evaluation as a discipline. Golden datasets, regression suites, A/B frameworks, hallucination and quality metrics, latency/quality/cost trade-off dashboards. If it isn't measured, it didn't improve.
5. Model economics. Model selection and routing, distillation, caching, and quantization decisions that cut inference cost without degrading quality — with numbers on both sides.

Hard requirements — verifiable, and we verify

• You have shipped LLM-powered features to production used by real paying users — not demos, notebooks, or hackathon projects — and can name the quality metric each feature moved and by how much.
• You have fine-tuned or preference-tuned an open-weight model (LoRA/QLoRA/full FT, DPO/ORPO or similar) for a production task, and can explain why fine-tuning beat prompting for that case — including a time it didn't.
• You have built an eval pipeline from scratch: dataset construction, metric design, automated scoring (including LLM-as-judge with its failure modes), and regression gating in CI.
• You have owned a retrieval system in production and improved a retrieval-quality metric (recall@k, MRR, downstream answer accuracy) with numbers you can state from memory.
• Strong Python; solid grasp of transformer internals (attention, KV cache, sampling) at the depth needed to debug generation behavior, not recite it.
• 4+ years in ML/AI engineering, with at least 1–2 years hands-on with LLMs in production.

Strong signal: low-latency or streaming inference experience · speech/audio ML · reinforcement learning from real-world feedback signals · small-model distillation · experience being the first/only ML hire.

Preferred education: IIT, Punjab Engineering College, Graphic Era, BITS, IIIT, IISc