AI Project Manager, Loop Engineer

Role DescriptionA NEW KIND OF PROJECT MANAGERBetaCraft builds software with a small human team and a fleet of AI coding agents. The agents write most of the code. What they cannot do is the thing this role is for: take a half-formed sentence from a domain expert, turn it into something buildable, get it built, prove it works, and make sure the same mistake never needs fixing twice.We call that loop engineering. A loop is: a stakeholder says something → it becomes a plan in plain language → a human approves it → agents build it in isolation → a gate checks it against evidence → a pull request opens → the stakeholder sees it working → and whatever was learned becomes a rule the next loop inherits. Loop engineering is designing that so a human touches it only where judgment matters, and it keeps closing when you walk away.We are hiring two to three senior people to own loops end to end. You will not groom a board or write a weekly deck. You will sit in the middle of a domain expert, a client's engineering lead, a second developer, an internal control plane, and six to eight parallel AI build lanes — and you will be the reason all of it converges.THE BASICSIn-office role, based in Pune.Full-time, Senior.STAKEHOLDER-DRIVEN DEVELOPMENT, AS WE PRACTISE IT— The stakeholder's words are the spec. Old documents are history, never authority. The hierarchy is: built code, then the client's primary sources, then our own specs.— Every request becomes a tracked story in our system. Every story gets a plan a non-technical person can approve. Approval is the only merge signal.— Thinking is front-loaded. We measure success by how often an approved story ships correctly with zero human touches after approval, and we always show that number beside its guard: how much came back within a week.— "Done" is a claim. Evidence is a real-environment walk, a screenshot, a number traced to its source, green CI on the final commit. You reject the rest, politely, every time.HUMANS ARE THE BOTTLENECK. WE DESIGN AROUND IT.Agents build in minutes. Humans answer in hours, or the next morning, or after the weekend. Every loop we run is shaped by that one fact, and this role owns the shape.— Thinking is front-loaded. The plan is argued before approval, so nothing waits on a human after it.— Park and continue. When a loop hits a question only a human can answer, it parks that item with a default and keeps moving. Nothing blocks on a ruling.— Decide the rest yourself. Only the genuine 50-50 calls reach the expert. Everything else you decide, record, and stand behind.— Primaries before people. A question goes to the expert only after you have proven the answer is not already in a primary source, an earlier ruling, or an old email. Then you batch, ask once, and store once.— The nightly loop runs alone and reports only. A regression walk runs every night; humans read one morning report and rule on the few parked items. That report is your control surface.— Two pipeline tiers. Small changes take the fast path; money, schema and cross-surface changes take the full path. You pick the tier, and a human sees it only when urgency and tier conflict.WHAT YOU WILL ACTUALLY DO— Sit with the domain expert. Hear a rule, restate it as a testable build order, ask "does that make sense?", and write it down where the next loop will find it.— Decide what to ask and what to build. If a primary source answers the question, you do not ask the expert. If you must ask, you batch, you ask once, and you store the answer once. The expert is never the bottleneck.— Design loops. Intake, plan, approve, build, gate, verify, record. You pick the pipeline tier per change, you write the impact brief before dispatch (what else this touches, what could break, who must say go), and you decide when a human is needed and when they are not.— Build harnesses. When something slips, you do not just fix it. You build the guard: a deterministic check, a parity test, a regression walk that runs every night and reports only, a morning report that is your control surface. Sweeps discover; guards lock. Every rule carries the incident that paid for it.— Run the agent fleet. Know when a lane is stale, when to kill it, when context is the bottleneck, when to say "don't build, just answer". Escalate only the genuine 50-50 calls; decide the rest yourself.— Hold the boundaries. The client's engineers own their code; you flag, you never touch. A second developer's work is gated before it merges. Merges are sequenced with an order and a parallel-safety note, never a bare list.— Communicate like a colleague. Firm, factual, collaborative, no blame. Every number you send is backed by a ledger you keep.WHO DOES WELL HERE— Problem-solving is your reflex, not a skill you list. You reduce a messy report to the actual failing mechanism before anyone touches it.— You read fast and completely. A long regulatory guide, a two-hundred-row issue log and a forty-file diff in one morning, holding the contradictions between them in your head.— You can read a code diff. You do not need to write code. You do need to notice that a fix landed in the wrong layer and ask why.— You see every bug as a class, and you also know when a class-wide sweep is too risky to run mid-acceptance.— You have product taste. You know a random banner when you see one, and you think about context before you delete an info icon.— You are calm under a date that does not move, and you would rather build the guard tonight than explain the same miss tomorrow.STRONG SIGNAL— You have been the single point of contact between a domain expert and an engineering team, and things shipped.— You have run or supervised AI coding agents on real work and have opinions about where they fail.— You have written a process rule after an incident that a team still follows.— Delivery in a domain where a wrong number costs money: tax, payments, insurance, compliance, analytics.NOT REQUIRED— Writing production code.— Prior knowledge of any specific domain. You learn each one from its primary sources.You will struggle here if your instinct on a hard question is "I'll check with the team", if you count tickets moved rather than proofs produced, or if telling a senior engineer that a green build is not evidence makes you uncomfortable.QualificationsProven experience managing complex engineering or software implementation projects, in technology-driven environments.Bachelor’s degree in Engineering, Computer Science, Project Management, or a related field.HOW WE HIREWork samples only. No behavioural interviews. About five hours over two weeks.Design the loop (take-home, 90 min): a stakeholder scenario with an expert, a client engineer and an agent fleet. Draw the loop, name where a human must touch it, and name the first three guards you would build.The "done" report (live, 45 min): an agent's confident pull request with passing tests. Say what evidence you would demand and what you would say to the agent.The expert relay (live, 45 min): a call transcript. Sort every item into build now, ask first, already answered, out of scope. Draft the one message you send back.Incident to guard (live, 30 min): a real post-mortem. Write the rule, then the mechanical check that enforces it.Conversation with Leadership Team (60 min): we show you our own process review, cons included, and ask where you think we are wrong.Apply with a short note and one thing you wrote that other people still use: a process rule, a post-mortem, a decision log, a runbook.To be considered for this opportunity, please complete our short application form: https://docs.google.com/forms/d/e/1FAIpQLSf3Kp0vjf-8TpgC3SCctw1p4MMO2HZMEgVUcZTxYpwZnyNOvg/viewform