Soraban

AI tax workflow platform for accounting firms

About Soraban

Soraban builds AI software for accounting firms. Firms use us to collect, chase, and extract client tax documents — and then to get that data into the tax software they already run, without anyone keying it in by hand. That last part is Connect, and it's the most important product we ship.

About Connect

Connect takes the data we extract from a client's W-2s, 1099s, K-1s, and brokerage statements and writes it directly into the firm's tax software — UltraTax, Lacerte, Drake, and CCH Axcess today, with more on the roadmap.

There's no clean API waiting on the other side. Most of this software is Windows desktop software, often running inside a hosting provider's virtual desktop. So Connect is a real distributed system: a Python service on AWS, a per-firm Windows worker fleet, a customer-installed agent, UI automation driving software that was never designed to be driven, and a QA layer that has to be right at the level of which box on which form.

It works, at real scale, in the busiest weeks of tax season. It is also demanding: exports have to land within a business day, workers have to stay connected, and a missed mapping can be very problematic to fix.

About the role

We're hiring a Connect Lead to own Connect end to end: the architecture, the reliability, the daily operations, and the roadmap.

This is a player-coach role. You'll set direction and run the function, and you'll also open the codebase, read the worker logs, and review the PR. We've found that people who only manage this from above lose contact with the actual problems fast — the failure modes live in screen resolutions, session states, vendor auth flows, and form field mappings, and you can't prioritize what you can't see.

You'll work directly with the CEO and the engineering leadership team. You'll have engineers building the platform, an ops team running the daily queues, and a QA function checking accuracy — your job is to make that whole system work, and to make it need less heroics every quarter.

What you'll own

• Reliability and the SLA. Connect has a one-business-day export commitment, and tax season doesn't respect business hours. You'll own uptime, latency, and the alerting and on-call model that supports it — including turning today's noisy alert streams into signals someone can actually act on.
• The architecture. Connect spans a Python/FastAPI service on ECS, a Rails/Postgres monolith, a fleet of per-tenant Windows EC2 workers built from versioned AMIs, and a customer-installed Windows agent. You'll make the calls on where logic belongs, how the fleet is provisioned and upgraded, and how we cut the cost and fragility of running hundreds of dedicated Windows machines.
• Data-entry accuracy. Every export is a claim about someone's tax return. You'll own the QA program — mapping correctness, prior-year overrides, form-level codes — and push it from manual per-firm review toward automated regression testing that scales to every new form and every new tax year.
• Operations, and automating them away. There's a daily queue of exports to shepherd and workers to reconnect. You'll run that function well and attack it: the goal is not a bigger ops team, it's fewer things that need a human.
• Onboarding new firms. Getting a firm live means credentials, MFA, hosting-provider security reviews, agent installs, and a test export. You'll own the time-to-first-export metric and the process behind it, including partner relationships with hosting providers.
• The roadmap. New tax software, new form coverage, entity returns, prior-year data matching, deeper API integrations. You'll decide what we build next and defend the decision with usage data.
• Communication. Sales, CS, and support all depend on Connect and can't read the logs. You'll set status definitions, escalation paths, and documentation that let the rest of the company answer customer questions without pinging an engineer.

What we're looking for

• 6+ years building and running production software, with at least a couple of them leading — a team, a platform, or a product area you were accountable for.
• Hands-on, still. You can read and write Python, review a PR on a system you didn't build, and reason about a Postgres query plan. You don't need to be the top committer, but you should be a credible one.
• Real infrastructure depth on AWS. EC2, autoscaling groups, AMI and launch-template lifecycles, ECS/Fargate, Terraform, Secrets Manager, CloudWatch. You've operated a fleet, not just deployed to one.
• You've owned reliability before. Alerting, on-call rotations, incident response, postmortems, and the unglamorous work of driving down alert noise until the pages mean something.
• Judgment about automation vs. headcount. You instinctively ask what fraction of a manual process can be removed before asking who else can do it.
• Excellent communication. Concise written updates, clear status definitions, honest escalation. Half this job is making a complicated system legible to people who will never see it.
• Goal-oriented and self-directing. You'll be handed a domain, not a backlog.

Strong pluses

• Windows automation experience — UI Automation, image-based RPA, RDP/DCV sessions, DPI and resolution handling, Windows service and installer packaging. This is unusual experience and it maps almost one-to-one onto our hardest problems.
• Tax or accounting domain knowledge. If you know what a K-1 box 17 code is, or have worked at or with a CPA firm, you'll move much faster.
• Integrations with legacy or hostile third-party systems — screen-scraping, unofficial APIs, vendor auth flows, systems that change without warning.
• Vendor and partner management — security reviews, hosting providers, software vendors.
• Experience in a seasonal business, where load and stakes spike predictably and the off-season is when you fix things.

What success looks like

In 90 days you know the system cold, you've been on call, you've shipped something that removed a recurring manual step, and CS knows who to ask and what the answer means.

By the end of your first year the export SLA is met without heroics, accuracy is verified by automated regression testing rather than spot checks, new firms go live in days without engineering involvement, and adding support for a new tax software is a project with a plan — not an expedition.

1. Intro call with the hiring manager (30 min)

2. Technical deep dive — a system you've built and operated (60 min)

3. Practical exercise: debug and design against a real Connect failure mode (90 min)

4. Cross-functional conversations with engineering, CS, and ops (2 × 45 min)

5. Final conversation with the CEO