Product Engineer (Senior) — VenXR

Venxr; Pune District; On-site; INR 3,000,000 - 5,500,000; Full time

Benefits offered by this job

Health cover

Team offsites

Flexible culture

Job summary

Venxr in Pune is seeking a Senior Product Engineer who builds with LLMs in production, owning features end-to-end from idea to live service. You'll work on campaign planning, model-in-the-loop solutions, and architecture decisions that influence how products behave at scale.

You'll collaborate across frontend (React), backend (Python/Django), and AI tooling, with a focus on orchestration, evaluation, and robust guardrails.

Qualifications

• Experience shipping an LLM system in production.
• 3–5 years shipping product.
• Strong debugging skills and ability to operate with ambiguity.
• Portfolio, GitHub, or project links to demonstrate work.

Responsibilities

• Build features end-to-end from idea to production.
• Work with model-in-the-loop systems and evaluations.
• Own architectural decisions and performance budgets.
• Collaborate across frontend, backend, and AI tooling teams.

Skills

React.js

Python

LLM systems

LangGraph

LangChain

Django

Orchestration

System design

Debugging

Technical leadership

Education

Bachelor's degree

Tools

Docker

CI/CD

PostgreSQL

Kubernetes

LangSmith

Job description

Product Engineer (Senior) — Venxr

Venxr builds AdsGPT: brands plan, run and optimise their ad campaigns by talking to it. We're hiring a Senior Product Engineer who builds with LLMs rather than around them — and can tell whether what they built actually works.

What You'll Build

Features, end to end — from a half-formed idea to production, and you own what happens after it ships. Most of what you build will have a model in the loop: campaign planning that reasons about a brand's goal, interfaces that assemble themselves from whatever the model surfaced, agents calling real tools against real ad platforms. At this level you also own the shape of things: the architecture calls that are expensive to reverse, the evaluation discipline that tells us whether an AI feature actually works, and the bar everyone else codes to.

Tech Stack

• Frontend: React.js
• Backend: Python, Django
• AI: Claude and other frontier models, LangChain / LangGraph, MCP

AI-native, and we mean it literally

Most job ads get this wrong, so plainly: using Claude to write code is not what we mean. We mean LLMs as the material the product is built from. You'll need working knowledge of:

• Fundamentals — tokens, context windows, cost, latency, structured outputs, why a confident model is often wrong
• Orchestration — LangGraph or equivalent: state, nodes, conditional routing, persistence, interrupt. Not prompts in a for loop
• Multi-agent patterns — when splitting work across agents helps, and the commoner case where it just multiplies failure modes
• Context engineering — what the model sees at each step: retrieval, compaction, selective injection. Where most of the quality lives
• Memory management — multi-turn state without the window or the bill growing unbounded
• MCP — connecting models to tools and APIs
• Evaluation — golden sets, regression gates, and knowing the true quality of a system that answers differently every time. Harder than building it, and where most teams quietly give up
• Guardrails and injection — user-controlled text flows into your prompts. That's a threat surface, not a footnote
• Cost and latency budgets — token spend per node, where the p95 actually goes, what's worth caching
• Observability — tracing, prompt versioning, and being able to debug a wrong answer at 2am from traces alone
• Knowing when not to — the cases where deterministic code is correct and a model is a liability

What we want is someone who has owned an LLM system in production, with real users on it, and can tell us honestly where it was weak and why.

What You Need

• Three to five years shipping product
• React.js, Python with Django — or FastAPI plus the obvious ability to get to Django in a fortnight
• An LLM system you owned in production with real users: the prompts, the orchestration, the failure modes, and the evaluation that proved it worked.
• An orchestration framework used in anger — LangGraph or equivalent, including the parts that broke
• REST APIs, SQL or NoSQL, and system design you can defend under disagreement
• Architecture judgement: you can name what you rejected and why, not just what you chose
• Technical leadership — code review, mentoring, raising the bar without being asked to
• Strong debugging, and comfort with ambiguity and almost no hand-holding

Your work speaks louder than your degree. Share links to your GitHub, portfolio, or projects you’ve built/shipped. That’s what we’ll look at first.

What Helps

• Conversational interfaces, and multi-turn state across channels
• LLM evaluation and monitoring — Ragas, LangSmith, Langfuse, or your own harness
• Working knowledge of Generative UI.
• AdTech, MarTech domain knowledge is a plus.
• Cloud infra, Docker, CI/CD
• Explainable or compliance-aware AI in a regulated domain
• Open-source work or writing we can read

How We Work

• Customer obsession — curiosity about what your code does to someone's day
• Ownership — chase blockers rather than wait on them
• Stakeholder sense — know when to pull in Product, Design or Business
• High standards — hold your work to a bar nobody is checking
• Disagree and commit — argue the call hard, then back it fully
• Simplify relentlessly — the simplest thing that works beats the cleverest thing that works
• Solidarity — show up for teammates unasked
• Vitality — energy that holds up in month six

Two of those carry extra weight at this level. Disagree and commit, because you'll be in the room where the calls get made. High standards, because yours become everyone's.

In practice: ship fast, test in production more than is strictly comfortable, and sit close enough to the founders that "I'll ask Product" means turning around.

The process

• Tech screening round — 30 minutes, mutual.
• Take-home — a small orchestrated LLM system, plus the evidence it works. Over half the marks are for judgement rather than completeness, and a meaningful share for telling us where your own system is weak. Using AI is expected; being unable to defend what it wrote is disqualifying.
• Interview round 1 — we walk through your submission and argue about it.
• Interview round 2 –Technical + Managerial

Baner, Pune. Health cover, team offsites, and a culture that's flexible in the ways that matter rather than the ways that photograph well.

Skills: langchain,mcp,system design,langgraph,django,react.js,orchestration,docker,ci/cd,langsmith