Python / LLM EngineerMid-Senior

Narix Labs; Jaipur; On-site; INR 3,000,000 - 5,000,000; Full time

Job summary

Narix Labs in Jaipur, India is hiring a Python / LLM Engineer to own the AI layer of Bispy, a live coaching platform grounded in behavioural-science methodology. You will manage prompts, context assembly, retrieval over Qdrant, and evaluation that keeps the system honest.

You’ll work end-to-end on the LLM coaching layer, tune vector stores, and build an explainable assessment pipeline in Python. The role emphasizes rigorous evaluation, data boundaries, and collaboration with product and client

Qualifications

• 3+ years of production Python
• 1+ year building LLM-backed features that shipped
• Hands-on retrieval/RAG experience with vector stores
• Familiar with multiple AI APIs (Claude, GPT, Gemini) and choosing empirically
• Strong written communication and documentation

Responsibilities

• Own the LLM coaching layer end to end: prompts, context assembly, retrieval, fallbacks, and cost control
• Build and tune retrieval over Qdrant: chunking, embeddings, hybrid search, and relevance measurement
• Maintain evaluation as a first-class system: golden datasets, automated scoring, regression gates on changes
• Develop the assessment scoring/ML pipeline in Python, with explainable results for non-engineers
• Keep data boundaries: sensitive assessment data handled with minimum-necessary treatment

Skills

Python

LLM features

Retrieval/RAG

Embeddings

Evaluation systems

API integration

Communication

Tools

Qdrant

PostgreSQL

Job description

Engineering

Python / LLM Engineer

Mid-Senior

Own the AI layer of a live coaching platform: an LLM assistant grounded in a real behavioural-science methodology, retrieval over Qdrant, and the evaluation discipline that keeps it honest.

Jaipur, India

All roles are based in Jaipur, India - on-site only, not remote.

The role

You’ll own the AI capability of a production platform: Bispy, an LLM coaching assistant that must stay inside a client’s behavioural-science methodology - an LLM freelancing generic advice would undermine the intellectual property the product is built on. The work is the unglamorous, decisive part of applied AI: retrieval quality over Qdrant, prompt and context engineering, evaluation suites that catch regressions before users do, and the scoring/ML pipeline behind assessments.

Role details

What you’ll do

• Own the LLM coaching layer end to end: prompts, context assembly, retrieval, fallbacks, and cost control
• Build and tune retrieval over Qdrant: chunking, embeddings, hybrid search, and relevance you can measure
• Maintain evaluation as a first-class system - golden datasets, automated scoring, regression gates on every change
• Develop the assessment scoring/ML pipeline in Python, with results that are explainable to non-engineers
• Keep data boundaries structural: sensitive assessment data gets the minimum-necessary treatment, always

What you bring

• 3+ years of production Python, with 1+ year building LLM-backed features that shipped
• Hands-on retrieval/RAG experience - vector stores (Qdrant or similar), embeddings, and the failure modes of both
• You treat evaluation as engineering, not vibes: you can describe how you’d catch a quality regression automatically
• Comfort with the full API landscape (Claude, GPT, Gemini, open-weights) and choosing empirically
• Strong written communication - model behavior gets documented, not remembered

Nice to have

• Classical ML fundamentals (scikit-learn-level) for scoring models beyond the LLM
• Experience constraining LLMs to a domain methodology or brand voice with measurable adherence
• Voice or real-time AI exposure

You’ll work with

PythonLLM APIsQdrantRAGEvalsPostgreSQL

What you get

The deal, plainly

• Senior-led delivery: your work is reviewed by people who’ve shipped production systems, and you see how they think
• An AI-native workflow - modern AI tooling is standard practice here, with human review as the quality bar
• Real products in production: client work that ships weekly and gets used, not internal demos
• Async-first, writing-heavy culture on IST, with deliberate US/EU overlap instead of late-night calls
• Competitive compensation, reviewed on outcomes