Software that talks to a million people a day. It has to work on the first call.
There is no staging environment for a conversation. When our AI picks up, someone with a real overdue loan or a lapsing policy is on the other end, in one of 10+ languages, on a line that crackles. That constraint is the job, and it is why the work here is unusually concrete.
Small company, large production surface.
Small enough that your work is identifiable in the product.
Live traffic across enterprise books, every day, not a pilot number.
Running revenue-critical conversations on the platform.
Profitable and bootstrapped. We have never needed a runway conversation.
Most AI work stops at the demo. Ours starts after it.
Latency is a design input, not a metric you tune later
A person hears a 400ms gap as hesitation and a 900ms gap as a broken line. Everything upstream of the voice, retrieval, reasoning, tool calls, has to fit inside a budget you can feel.
We picked the hardest place to build on purpose
10+ languages, mid-sentence code-switching, adversarial callers who have heard every script, and telephony older than some of the team. Anything that survives here is over-engineered for anywhere else.
We are not chasing 100% automation
AI resolves the bulk; a person closes the settlement, the save, the exception. We treat the AI-to-human handoff as a first-class product, which is a harder and more interesting problem than the automation rate.
Five places the work sits.
Titles vary. What does not vary is that every one of these roles is measured on whether a real conversation ended in a real outcome.
Conversation models, retrieval, evaluation
Grounded answers over a client's actual policy and account data, with an evaluation harness that runs on production audio rather than a curated set.
- Speech pipeline tuning for Indian audio conditions
- Retrieval and grounding against client systems of record
- Automated scoring across every call, not a QA sample
Runtime, telephony, integrations
The dialler, the call runtime, the CRM and core-system connectors, and the control plane that lets an operations team change behaviour without a release.
- Real-time media and telephony at scale
- Integration work against legacy enterprise stacks
- Observability for systems where the user is on the line right now
Making the outcome land in a live book
The people who take a signed programme and make it hit the number, working alongside the client's own floor. Closest to the truth in the company.
- Programme design against a client's real book
- Baseline agreement, measurement and outcome realisation
- Handoff design with the client's human agents
Selling an outcome, not a licence
Enterprise BFSI and telecom sales cycles, and the positioning, research and content that makes the category legible before a rep ever calls.
- Enterprise sales into banks, NBFCs, insurers and telcos
- Product marketing and original research
- Design and front-end for everything customer-facing
Four steps. No unpaid project marathons.
A real conversation first
Not a screen. You will speak to someone who does the work, and you should leave the call knowing what the job actually is. Ask us what is broken. We will tell you.
A problem from our actual domain
Time-boxed and drawn from something we genuinely faced, so it tells you as much about us as it tells us about you. We do not ask for free work on live tickets.
Meet the people you would sit with
Including at least one person outside your function, because delivery, engineering and revenue are not separable here.
A decision, quickly, either way
We aim to close a loop within two weeks of the first call. We would rather tell you no in a week than leave you waiting for a month. If we say no, you will hear why.
What we are not going to pretend
- Some of this work is unglamorous. Accents, line noise, retry logic, DND windows, a regulator's calling-hour rule. The interesting problems here are wearing boring clothes.
- We are a real company with real clients, which means real deadlines. Enterprise programmes have go-live dates, and some weeks bend around them.
- Bootstrapped cuts both ways. No runway anxiety and no theatre, but also no infinite budget. We make trade-offs out loud.
- We are in a category that is being defined right now. That means ambiguity is part of the job description, not a temporary condition.
- If you want to work only on models, this is the wrong place. The value here is in the last mile, where the model meets a phone line, a compliance rule and a person who wants their problem solved.
The live list, and the route if nothing fits.
We are not running an open list on this page yet. If your work fits one of the five areas above, write to us directly and tell us which one, and what you would want to own in the first ninety days.
If you are not looking but want to understand what we are building, the platform pages are the honest version, written for buyers and equally readable as a description of the engineering problem. Start with the handoff.
The best way to understand the job is to hear the product work.
Put your own number in and our AI will call you. It is the same thing a bank's customer hears, and it is the clearest interview prep we can offer.