About Donna

Donna is the time-to-cash platform for law firms. This role is for our AI time tracking product for lawyers: we watch how a lawyer actually works across their whole practice and turn it into accurate, billable docket entries with written narratives. Timekeeping is the task every lawyer hates and most do badly, and it sits directly on top of firm revenue, so getting it right is worth a great deal to them. We are live, we have paying users, and the core bet is proven: lawyers want this.

That puts us at the best moment in a product's life. Existence is settled; now it is about depth. Getting from "this is impressively close" to "I trust this with my billing" is a genuinely hard systems problem, and owning it end to end is the whole of this role.

The problem you would work on A lawyer's day is not a series of discrete tasks. It is one continuous, interleaved stream: a motion drafted across a word processor, three browser tabs, a practice management system, and two interruptions, all inside forty minutes. A human reconstructing that day knows instantly that it was one piece of work for one client.

Teaching software to see it that way is the core intellectual challenge of our product, and it is a genuinely new problem, not a solved one you would be re-implementing.

The interesting part is that this is not a model-intelligence problem. It is an evidence problem: what signal you capture, how faithfully you preserve it, how you assemble it into a picture of real work, and how you know whether the picture is right. Knowing which client a piece of work belongs to, and knowing when you are not sure, is the difference between software a lawyer tolerates and software they trust with their revenue. Clearing that bar is the win.

You would work closely with the founders to help solve this problem for lawyers. The work is measurement, evidence integrity, and systems architecture, not prompt engineering: the model layer is not our bottleneck, and we have the evidence to say so, so your effort goes into the parts of the system that actually move accuracy. It starts as measurement and architecture and grows into technical ownership of output quality as the product expands.

You would be our second engineer on time tracking, with a real path to a permanent or founding-engineer role if it goes well.

We ship daily and run fast experiments, and we are also building rigorous measurement. Those sound like opposites; reconciling them is the whole job. Velocity with measurement compounds, and velocity without it is just motion. We ship fast, and we want to ship fast and know it is getting better. The interesting problem here is not intelligence. It is knowing, provably, what your system got right, and few products give you a cleaner shot at solving it.

What you would own

Product:





- The input layer and capture architecture. Everything downstream is bounded by the fidelity of what we observe. This is the foundation, and we believe it is where the highest leverage sits, which makes it the best place to start.
- Interpreting cross-application workflows. Recognising one continuous piece of legal work as it moves across applications, documents, browsers, and meetings, and keeping it whole instead of shattering it into fragments or blending two clients together. This is the heart of the product.
- Accurate context building. Every matter accumulates a working understanding over time. Getting that to compound correctly, learning from confirmed facts with clear provenance for everything we believe, is one of the most consequential design problems we have and one of the most rewarding to get right.
- Work objects. Documents, meetings, and email are each becoming first class things the system reasons about rather than undifferentiated activity. You would shape what that means.
- Knowing when we are right. Defining what accuracy even means for a docket entry, and making it measurable, is a product question as much as a technical one, and it is wide open for you to define.

Technology:

- The evaluation harness. We have real captured days and a large body of expert-corrected ground truth already in hand. Turning that into a rigorous, gating eval loop, so that every change to the product is a measured change, is the single highest-leverage thing an engineer can build here, and the raw material is waiting for you.
- The desktop application. A signed, auto-updating Electron app on macOS and Windows with a local Python pipeline, plus a hosted virtual desktop pilot. A real cross-platform surface with real, interesting constraints.
- Third party integrations. Practice management systems, document stores, calendars, and email. The lawyer's real workflow lives in other people's software, and meeting it there is a large part of accuracy and a large part of the moat.
- The pipeline itself. Python, local-first classification, sync to a regional cloud backend. Dense, thoughtfully built, and ready for someone to raise the bar on.

How we work: We are a lean team and we move fast, and we like it that way. This is not a 9-to-5. We show up when it is required, we make our own calls about what matters, and we trust you to do the same.

Concretely, day to day:

- We run fast experiments. An idea gets built and tested in days,



not sprints. Code is quick and cheap to write now, so the bottleneck is not typing, it is knowing what to build and knowing whether it worked. We use AI tooling aggressively and you would too.
- We ship constantly. Multiple production releases a week, pushing toward daily. A short, satisfying path from decision to users.
- Turnarounds are quick. If something is broken for a lawyer, it is broken today, and we fix it today.
- You own outcomes, not tasks. Nobody will tell you what to do on Tuesday. We will tell you what has to be true in a month, and get out of your way.

If you want a clearly scoped backlog and predictable hours, we are honestly not the right fit. If you have been the person who quietly made the whole thing work at 11pm on a Thursday because it mattered, and enjoyed it, you will recognise us.

What we are looking for

Required:

- You have run an evaluation harness against a real production system, not a benchmark. You have had to answer "did this change actually help" with a number, in front of someone who disagreed, and been right.
- Strong Python, and comfort across a full product surface (desktop, backend, data).
- Instincts for data provenance and feedback loops. You think carefully about where a system's beliefs come from and whether they can be trusted.
- Judgment about failure modes. The question that excites you is not only "is it accurate" but "what does it do when it does not know."
- You move fast without breaking trust. We ship daily and we bill lawyers, and holding both of those at once is the craft of the job.
- Directness with founders. We want someone who will tell us we are working on the wrong layer.

Useful but not required:

- Desktop capture experience (Electron, OS accessibility APIs, OCR).
- Legal tech, billing, compliance, or any domain where a wrong output has real consequences for a real person.

If you have read this far: message Saumya Banker on LinkedIn with the words "read the jd" + 1 sentence on what you would measure first. It is the fastest route to a conversation.

Practicalities

- Stack: Python (FastAPI), TypeScript/React, Electron, Supabase, Firebase. Monorepo, pnpm.
- Shape: Contract, 3 months initially, with a strong bias to extend. Full-time. Remote. Start ASAP.
- Interview process: an intro call with our CEO, a technical round with our CTO, a conversation with our CPO, and a final culture-fit round. All cofounders, and we move fast between stages.
- First week: once you are hired, we hand you our own architectural assessment and the harness, and let you loose. What we want to see is how you decide what matters most and whether you can prove you moved it. Your week-one deliverable is a production-ready build addressing the biggest problem, as you diagnose it.

📌 Senior Software Engineer, AI Systems (India)
🏢 Donna AI
📍 India