ASTRM Intelligence Labs is building the software and autonomous systems that armed forces will fight with. We are building two halves of one system. The first turns the flood of raw intelligence, from imagery and signals to open source and sensor feeds, into decisions a commander can act on in minutes instead of days. The second puts autonomy on the physical assets that act on those decisions: drones, ground robots, and the coordination layer that makes a heterogeneous team of them behave like one. Every line of it is built in house, on infrastructure we own. Sovereign capability is not a marketing line for us. It is an engineering constraint that shapes every decision we make, from the models we train to the hardware we solder. We are a small and nimble team. Our work is already in the hands of users in uniform. What we ship this year is what gets used at the edge next year.
The problem
Our software has to run where the network does not. On a rack in a formation headquarters with no route to the internet. On a ruggedised box in a forward location that gets power for six hours a day. On compute strapped to an airframe. The same mission software, the same models, the same data, across three radically different environments, all of them disconnected, all of them unforgiving of an operator having to debug a deployment at 0300.
Somebody has to own the substrate that makes this possible. That is this role.
You will build the platform that every other engineer at ASTRM deploys onto: the compute fabric, the data plane, the model-serving layer, the build and release system, and the store-and-forward machinery that keeps an edge node and a headquarters node coherent across a link that is intermittent by design.
What you'll own
- The deployment substrate. Reproducible, air-gapped installs of the full ASTRM stack onto customer hardware, from a single ruggedised node to a multi-GPU rack. Bootstrapping, orchestration, upgrades and rollback that a signals officer can execute from a runbook, offline, without us on a call.
- The data plane. Ingest, storage and retrieval across wildly heterogeneous data: imagery and video, time-series telemetry, signal captures, documents, entity graphs. Choosing and running the stores behind it (relational, time-series, vector, graph) and making them fast under a query load that spikes when things are happening.
- Model serving at the edge. GPU inference infrastructure for vision and language models on everything from a Blackwell rack down to an embedded module. Batching, quantisation, multi-model residency, keeping latency honest when memory is the binding constraint.
- Edge-to-HQ synchronisation. Conflict resolution, prioritised replication and graceful degradation across links that drop for hours. Deciding what a disconnected node is allowed to conclude on its own.
- Release engineering for classified environments. Build pipelines, artefact provenance, dependency vendoring and a CI story that works when the target network has never seen the public internet.
- Observability where you cannot SSH in. Telemetry, health and diagnostics designed to be read from a bundle a customer hands you on a drive three weeks later.
What we're looking for
- 3–8 years building and running production infrastructure across distributed systems, platform engineering, ML infrastructure or similar. Senior-leaning: you should have owned something significant, not just contributed to it.
- Deep comfort with Linux, containers and orchestration, and with the ugly parts underneath them: networking, storage, kernel-adjacent debugging, GPU drivers.
- Strong systems programming in at least one of Go, Rust, C++ or Python, and the judgement to know which of them a given problem deserves.
- Real experience running GPU workloads, including serving stacks such as vLLM, TensorRT-LLM or Triton, and an intuition for where memory and latency actually go.
- Hands-on with the data infrastructure (Postgres/PostGIS, time-series stores, vector databases, message queues) and opinions about when each is the wrong choice.
- A bias toward boring, legible, debuggable systems. Our operators cannot page you.
Strong signals
- You have shipped software into an air-gapped, on-premise or otherwise hostile deployment environment and lived with the consequences.
- Experience with embedded or edge compute: Jetson-class hardware, cross-compilation, thermal and power budgets.
- You have built the internal tooling that made a team of engineers meaningfully faster.
- Security engineering instincts around secrets, supply chain and hardening, plus the judgement to know what actually matters versus what is theatre.
- Prior work in defence, aerospace, industrial, medical or another domain where a bad deploy is not just an outage.
Why this role is unusually good
You will define the platform from close to zero, with real users and real hardware, and no legacy to inherit or excuse. Nobody will hand you a cloud provider's abstractions to hide behind. Within a year you will have built infrastructure that runs in places most engineers never get to see.
Location: Bengaluru. On-site. This is hands-on work with hardware, secure environments and each other, and it does not work remotely.
Compensation: Competitive cash, and meaningful equity. Early employees at ASTRM own a real share of what they build.
How to apply: Send your CV and a short note to [email protected] with the role in the subject line. Skip the cover letter. Instead, tell us about one thing you built end to end and what broke along the way. Links to code, papers, teardowns, build logs or flight footage are worth more to us than a polished résumé.
On experience: The ranges mentioned are a guide, not a gate. If you are well short of them and can show us work that says otherwise, apply anyway. We have hired that way before.