Build the thing that listens.
We are small, senior and deliberately unglamorous about it. The work is hard in specific ways: real-time constraints, regulated customers, and other people's hardware.
Five roles, all of them senior.
If none of these fit but you think you should be here, write to us anyway and tell us what you would work on. We have created roles for the right person before.
Own the real-time path: streaming ASR, diarisation, endpointing and the latency budget that makes a hint arrive inside the sentence rather than after it.
What you will do
- Ship low-latency streaming inference that holds up on a noisy PSTN line, not just clean audio.
- Push speech-to-speech translation latency below 400ms across eleven Indian languages.
- Make it all run on a customer’s hardware, sometimes without a GPU you would have chosen.
What we are looking for
- Deep experience with streaming ASR or TTS in production, not just fine-tuning.
- Comfort with quantisation, batching and the unglamorous work of making models fit.
- You have debugged an audio pipeline at 2am and enjoyed at least part of it.
Build the grounding layer: indexing, version suppression, ACL-aware retrieval, and citations precise enough that a compliance officer will accept them.
What you will do
- Make retrieval correct before making it clever — a wrong cited answer is our worst outcome.
- Handle document versioning so a superseded 2024 brochure can never answer a 2026 question.
- Honour document-level access control at query time, per user, without wrecking latency.
What we are looking for
- Production RAG experience, including the failure modes nobody blogs about.
- Strong evaluation instincts: you measure retrieval quality rather than vibe-check it.
- Interest in regulated domains where citation matters more than fluency.
Get ira running inside banks, insurers and telcos — their hardware, their change process, their security team asking hard questions.
What you will do
- Own installs end to end: sizing, networking, telephony integration, SSO, go-live.
- Sit with customer security teams while they inspect egress and read our logs.
- Bring what you learn back to product — you will be the reason things get easier.
What we are looking for
- Kubernetes, Linux and real networking knowledge, including air-gapped environments.
- Calm under a room full of stakeholders during a cutover.
- Willingness to travel to customer sites regularly.
Sell a technical product to regulated buyers who have been disappointed by AI vendors before, and who are right to be sceptical.
What you will do
- Run full cycles with BFSI and telco buyers, from first call to signed MSA.
- Work with security and legal early, because on-prem shortens those reviews rather than avoiding them.
- Say no to deals we should not take. We would rather lose a logo than a deployment.
What we are looking for
- Enterprise software sales experience in India, ideally BFSI.
- You can hold a technical conversation without a solutions engineer in the room.
- You treat procurement as part of the sale, not an obstacle after it.
Design an interface an agent reads while talking to a distressed customer. Every extra element costs them attention they do not have.
What you will do
- Design for peripheral vision and split attention, not for a portfolio screenshot.
- Work on the hint card, the compliance monitor and the AutoCoach scorecard.
- Sit in on live calls until you feel the pace yourself.
What we are looking for
- Strong systems thinking and craft in dense, information-heavy interfaces.
- Experience designing tools people use all day, under pressure.
- You prototype in code or work closely enough with engineers that it does not matter.
Worth knowing before you apply
You will meet customers
Engineers join deployment calls. Designers listen to live conversations. Nobody here builds for a persona they have never heard speak.
Hard constraints, on purpose
On-prem, no outbound calls, latency budgets in milliseconds. The constraints are the product, and working inside them is the interesting part.
Small team, real ownership
You will own a surface rather than a ticket queue, and you will be the person customers ask for by name when it matters.
Four conversations, about two weeks
Intro call — 30 minutes
What you have built, what you want to build next, and an honest description of the role from us.
Working session — 90 minutes
A real problem from our actual backlog. No whiteboard algorithm puzzles, no take-home that eats your weekend. We pay for take-home work if you would rather do it that way.
Team conversations — 2 × 45 minutes
The people you will work with daily, including someone from outside your discipline.
Decision within 5 days
We tell you either way, with reasons. If it is a no, you will know why.
We are an equal opportunity employer. If you need any adjustment to the process — timing, format, accessibility — tell us at the intro call and we will arrange it, no explanation required.
Not on the list?
Tell us what you would build here and why it should exist. careers@ira.ai reaches a founder.