Senior Software Engineer - Intelligence

Smart Joules; New Delhi; Vor Ort; INR 1.200.000 - 1.800.000; Vollzeit

Zusammenfassung

Smart Joules Pvt. Ltd. is seeking a Senior Software Engineer for its Intelligence team in Delhi.

You will own significant parts of the stack end-to-end, from machine learning training and serving to AI agent runtimes and data platforms. You’ll work closely with customers and campus teams, shaping what gets built and how it behaves in production. You will join a small team with broad ownership, comfortable with ambiguity, and ready to take on technically challenging edge and cloud integration

Qualifikationen

• 4–5 years of building and running production software.
• Strong Python and depth in at least one of Node.js, a front-end framework, or an ML serving stack.
• Strong statistical foundations and data exploration skills.
• Experience with ML systems in production, drift and retraining challenges.
• Clear written communication and ability to work with product managers or customers.

Aufgaben

• Own the design, implementation and production behavior of a subsystem, including failure modes.
• Take ML models from experiment to production — training, deployment, monitoring, retraining.
• Work on learned control policies governing real machinery.
• Build AI systems with retrieval and tool-calling agents grounded in a knowledge graph and telemetry.
• Debug across the stack from edge devices to APIs and UI seen by users.
• Present work to technical and non-technical audiences and write design docs.
• Review code and mentor junior engineers.

Kenntnisse

Python

JavaScript/TypeScript

ML in production

Data pipelines

Edge computing

Tools

AWS SageMaker

Airflow

PostgreSQL

Redis

Docker

Kubernetes

Jobbeschreibung

About the role:

As a Senior Software Engineer on the Intelligence team, you will own a significant part of that stack end to end. Depending on your background and interest, that could be the machine learning training and serving path, the AI agent runtime, the data platform, the learned control policies running on our edge devices, or the backend services that connect them.

Ownership here means more than writing the code. It means understanding how your subsystem fails, deciding what should not be built, explaining the trade-offs to the people who will live with them, and being someone others come to when a number looks wrong.

You will also work close to the customer. Our users are facility teams, energy managers and plant operators, and the fastest way to understand why something is not working is to watch one of them try to use it. Engineers here shape what gets built, not only how it gets built.

We are a small team with a large surface area, so you will have unusual latitude and unusual responsibility for the parts you take on.

Our tech stack

• Languages — Python (primary), JavaScript, TypeScript


• Machine learning — scikit-learn, XGBoost, LightGBM, Keras, pandas, numpy, AWS SageMaker


• Reinforcement learning and control — RL policies, classical PID, lookup-based control, running on constrained edge hardware


• AI systems — AWS Bedrock, agent loops, tool registries, retrieval-augmented generation, prompt caching, model tiering, grounding and verification, evaluation harnesses


• Data platform — Apache Airflow, InfluxDB, Timestream, PostgreSQL, Redis, DynamoDB, S3


• Knowledge graph — Amazon Neptune, Brick schema, RDF and SPARQL


• Observability — Grafana, Prometheus, Loki, Sentry, OpenTelemetry, CloudWatch


• Backend and front end — FastAPI, Node.js with streaming interfaces, Angular


• Edge — Python on ARM Linux, MQTT, local PostgreSQL, systemd


• Cloud and infrastructure — AWS, Docker, GitHub Actions

What you will do:

• Own the design, implementation and production behaviour of a subsystem, including its failure modes


• Take machine learning models from experiment to production — training pipelines, deployment, monitoring, retraining, and an honest assessment of when a prediction should not be acted upon


• Work on the learned control policies whose decisions become physical actions on real machinery


• Build and improve our AI systems: retrieval and tool-calling agents grounded in a knowledge graph and live telemetry, with evaluation frameworks that measure whether the answers are actually right


• Debug across the whole stack, from an edge device in a plant room through message transport, storage and APIs to what a user finally sees


• Partner with product managers and customers to define what gets built, and offer a different view, with reasoning, when a requirement does not hold up


• Present your work to technical and non-technical audiences alike, including customers


• Review code, write design documents others can act on, and mentor engineers earlier in their careers

You may be a good fit if you have:

• Around 4–5 years building and running production software, with strong Python and working depth in at least one of Node.js, a front-end framework, or an ML serving stack


• Machine learning in production — models that real decisions depended on, and first-hand experience of drift, retraining, and the gap between an offline metric and live behaviour. Depth in classical ML, deep learning and neural networks, or both


• Strong statistical foundations and the instinct to explore data before modelling it. Much of our work begins with messy sensor data, and knowing what a distribution, a correlation or an outlier is actually telling you decides whether the model that follows is worth anything


• LLM systems in production — a retrieval-augmented generation pipeline, an agent with tool use, a structured extraction system, or something similar, that real users depended on. Built on AWS Bedrock, LangChain, or direct provider APIs. We are as interested in how you evaluated it and how you prevented confident wrong answers as in what you built


• A working understanding of reinforcement learning — agents, states, actions, rewards, and why an RL policy behaves differently from a supervised model. We run RL in production, so hands‑on depth is a strong advantage


• Fluency with AI coding assistants such as Claude Code, Codex or Cursor as part of your daily work. A large share of first-draft code here is AI-assisted, so knowing how to review it, catch its errors and decide what to keep is a core skill


• Worked with data infrastructure at some depth — time‑series databases, caching or streaming systems — and formed opinions about staleness, consistency and back‑pressure


• Been responsible for something with real users and real downtime, and learned from it


• Experience collaborating with product managers or customers to decide what to build, not only to receive specifications


• Clear written communication. Design documents and post-mortems here are read widely

Nice to have:

• Open source contributions — a library you maintain, sustained contributions to a project, or work that others build on


• Reinforcement learning applied to a real physical system rather than a simulator alone


• Deep learning applied to time-series, sequence or control problems


• Experience with building automation, HVAC, or industrial control systems


• Knowledge graphs, ontologies, or semantic data modelling


• Running orchestration platforms such as Airflow at a scale where they became difficult


• Embedded or edge deployment on constrained hardware


• Having owned a product area rather than only a codebase