Find Work

Our search is keyword only. "Something warm for winter" returns nothing; "kurta" returns 400 results in no useful order. We want semantic search with sensible ranking, and — importantly — it must still nail exact matches. Semantic search that gets fuzzy about a specific SKU code someone pasted in is a regression, not an upgrade. We want relevance measured, not asserted. Tell us how you would build an evaluation set from our search logs.

Our roguelike generates levels with hand-written rules that produce a lot of layouts that are technically valid and no fun. We want a generator that can be tuned against measurable properties — path length, backtracking, encounter spacing — and a way to evaluate a generated level automatically so we can iterate without playing 200 of them. Unity, C# for the runtime, Python for the tooling side is fine. This is a design problem as much as a technical one. Please tell us how you would define "fun" as something measurable here.

We want to give members feedback on squat and deadlift form from their phone camera. This needs to run on-device — sending workout video to a server is both slow and something our members would object to. Latency has to be low enough to feel live. We are realistic about scope: three exercises, a handful of common faults each, and confident silence when it is not sure. We would rather it says nothing than tells someone their good form is wrong. Our physiotherapist will define the fault criteria. You build the detection.

We want a tool that turns "seven days in Kerala, two adults, mid-range, no early mornings" into a real itinerary built from our actual bookable inventory. The hard requirement: it must never invent a hotel, a price or a transfer that does not exist in our system. Every item in the output has to come from a tool call against our API. We have seen demos of this that hallucinate a beautiful trip nobody can book. That is worse than useless to us — it is a customer service problem. Our API is documented and ready.

We publish a weekly hour-long interview and cut short clips from it by hand, which takes an editor most of a day. We want transcription with speaker labels, then suggested clips — self-contained moments of 30-60 seconds — ranked so an editor reviews ten instead of scrubbing an hour. Handles Indian English and code-switching into Hindi mid-sentence, which most off-the-shelf transcription does badly. The editor stays in the loop. We are removing the scrubbing, not the judgement.

We work with about 400 contract farmers who send us photos of affected crops on WhatsApp, and our agronomist answers them one by one. We want a first-pass classifier for the twelve most common issues in the crops we source, with a confidence score that routes the uncertain ones to the agronomist. Photos are exactly what you would expect: phone cameras, poor light, sometimes out of focus. Trained on clean lab images this will fail, and we would rather you tell us that now. We have about 5,000 labelled photos from real submissions.

Our return rate is 14% and the reason codes are almost useless because 60% of returns are logged as "other". We want a dashboard that combines reason codes with what we can infer — size, category, price band, channel, whether the customer reordered a variant. The deliverable that matters is a view our category managers can use to decide which products to fix, reshoot or drop.

We spend across Meta, Google, YouTube and some offline, and platform-reported conversions add up to more than our actual sales. Everyone is claiming the same customer. We want a media mix model that gives us an independent view of contribution, with adstock and diminishing returns handled properly. Two and a half years of weekly spend and revenue, plus promotion calendar. We would rather have honest wide confidence intervals than a precise-looking number nobody should trust.

Our game runs at 60fps on a flagship and stutters badly on the ₹15,000 Android phones most of our players actually own. We want profiling on real mid-range devices, then the fixes: draw calls, GC pressure, texture memory, whatever the profiler actually says rather than what we assume. We have a device lab of five phones we can ship to you. Hourly, because we want you to stop when the returns flatten rather than pad a fixed scope.

Our infrastructure was clicked together in the AWS console over two years and only one person knows where everything is. We want it in Terraform, a deployment pipeline that runs tests and can roll back, and environments that actually match each other. We are aware this is partly an archaeology job. Part of the deliverable is a written account of what exists, including the things that turn out to be running for no reason. Production handles money, so we need a migration plan that does not involve a big-bang cutover.

We find out our data is wrong when someone notices a dashboard looks odd, usually days later. We want automated checks — freshness, row counts, nulls in columns that should never be null, referential integrity between marts — that alert before a human notices. The design constraint is alert fatigue. A check that cries wolf twice a week will be muted within a month, so thresholds need thought rather than defaults.

We check competitor package prices by hand once a week across about a dozen sites. It is slow and inconsistent. We want it collected automatically and normalised into a comparable structure — the hard part, since everyone bundles differently and "5N/6D Goa" means different things. Only public listing pages, respecting robots.txt and sensible rate limits. We are not asking anyone to work around access controls, and a proposal that offers to will not be considered.

Three hotel suppliers, three different APIs, three different ideas of what availability means. Our booking flow currently supports one. We want an adapter layer with a single internal interface, sensible caching, and graceful degradation — if one supplier is slow we show the others rather than hanging the page. The interesting problems here are timeouts, partial failures and reconciling inventory that disagrees. Please address those in your proposal; the happy path is the easy part.

We have about 9,000 reviews across our own site and three OTAs, and our understanding of them is an average star rating. We want them analysed by topic — food, transfers, guides, accommodation, value — with sentiment per topic per destination, so we know what to fix where. Mixed English and Hinglish, which trips up most out-of-the-box sentiment tools. The output should be a repeatable report, not a one-off deck.

Our operations team works out of the database with a SQL client, which is as alarming as it sounds. We want an internal admin tool: search, view, and a controlled set of edits, with roles so a support agent cannot do what only a manager should, and an audit log of every change. Not a beautiful product. It needs to be clear, fast and hard to do damage with. About 15 screens.

We process recorded field interviews in Hindi, Marathi and Telugu. Off-the-shelf transcription handles the Hindi acceptably and the other two badly. We want a pipeline with per-language handling, speaker diarisation, and output good enough that a human correcting it is faster than typing from scratch — which is the only bar that matters commercially. Recordings are field quality: outdoors, background noise, occasional overlapping speech. Please be honest about achievable accuracy per language. We would rather scope realistically.

Members check in on a paper register. We want a tablet at the door and a QR code in the member's app. Must work when the studio wifi drops, which it does — queue the check-ins locally and sync when it comes back. Duplicate check-ins on sync are the failure mode we want handled properly. Also needs a simple trainer view of who is in the room right now. Four studios, roughly 2,800 members.

We schedule about 160 classes a week across four studios and three trainer specialities, currently in a spreadsheet by one person who is very good at it and would like a holiday. Constraints: trainer availability and speciality, room capacity, member demand by time slot, and a minimum gap between a trainer's classes. We want a solver that produces a workable schedule and, when it cannot, says which constraint is the binding one. That last part is what makes it usable.

We ship three LLM-backed features and we find out we have made them worse when a customer tells us. We want a proper evaluation harness: a golden set, automated scoring where it can be automated, human review where it cannot, and it running in CI so a prompt change shows its effect before merge. We are more interested in someone who has learned what does not work here — LLM-as-judge without calibration, for instance — than in someone who has read about it.

Our collections team calls accounts in order of days overdue, which means they spend equal effort on someone who was going to pay anyway and someone who never will. We want a prioritisation model: who is most likely to pay if contacted now. Not a risk score — an action score. We have four years of contact attempts and outcomes. It has to be explainable enough that a team lead can override it and tell us why.