Find Work

We onboard vendors on the strength of an email and a bank detail, which our auditor has now noticed. We want a due diligence process proportionate to our size — not a bank's onboarding pack, but enough that we know who we are paying: entity verification, GST status, beneficial ownership for larger vendors, and a risk tiering. Deliverable is the process, the checklist, and the templates, plus a short note on what we should do about the 60 vendors already onboarded.

We are raising and expect technical due diligence in the next quarter. We would rather find our problems ourselves first. We want an experienced engineer to review architecture, security posture, dependency and licence position, and delivery process, and tell us plainly what a diligence team will flag. A blunt list is more useful to us than a diplomatic one. We have time to fix things now and will not later. Roughly 30-40 hours over a few weeks.

Travel is brutally seasonal and we take deposits months before we pay suppliers, which flatters our bank balance in a way that has caught us out twice. We want a 13-week rolling cash flow model that treats customer deposits as the liability they are, and a longer annual view for planning. It has to be maintainable by our finance manager, so Excel rather than something clever, and documented. Three years of history available.

We discount on instinct at end of season and we have no idea whether a 30% markdown sold stock we would have sold anyway at 10%. We want elasticity estimated per category, and a markdown schedule that clears stock by a target date at the least margin sacrificed. The difficulty is confounding — we discount when things are not selling, so naive regression will tell us discounts reduce sales. Please say how you would handle that; it is the main thing we will judge proposals on. Three years of daily price and units by SKU.

Our site is pretty and converts at 0.7%. We would rather it were plainer and converted at 3%. Scope is the core marketing pages, rebuilt in Webflow, with copy as part of the job — the copy is most of the problem. We want the reasoning made explicit: what you changed and what you expect it to do, so we can check afterwards rather than admire it. Set up so we can run tests ourselves after handover.

B2B SaaS, about 900 accounts. We want to predict which accounts are at risk 60 days out. We have built one of these before and it failed for a familiar reason: it predicted churn accurately and nobody knew what to do with the prediction. This time we want the model to surface which drivers moved for that account, so customer success has something to act on. And we want the loop closed — record which accounts were contacted and what happened, so in six months we can tell whether the intervention worked at all. Usage, billing and support data available.

By the time an account is 30 days overdue our options are limited. We want a signal 30 days before that. Inputs are repayment behaviour, transaction patterns in the linked account, and changes in usage. Output is a weekly risk movement per account — we care more about accounts that have deteriorated than about the absolute level. Regulated context, so interpretability is a requirement rather than a preference, and we need model documentation to match. Four years of history, roughly 18,000 accounts.

We get about 300 trial signups a month and convert 22%. Our sales team calls them in the order they arrive. We want a score so they call the likely converters first, built from what we know at signup plus early behaviour — did they book a class, did they attend it. Small data by ML standards, roughly 7,000 historical trials. We would rather have a simple model we understand than something overfitted to that. Deliverable includes the score wired into our CRM as a field.

We publish about 40 pieces a month and the variance in performance is enormous. Our editors have instincts about what will do well and we have never tested them. We want a model that estimates reach for a draft from its topic, headline, length, format and timing — and, more useful, tells an editor which lever matters most for this particular piece. We are aware this may turn out to be weakly predictive. If the honest answer after analysis is "the signal is mostly in the topic and everything else is noise", that is a valid deliverable and we will pay for it. Three years of published content and analytics.

We block hotel inventory months ahead and either sell out early at a bad price or hold rooms we cannot fill. We want demand forecasts by destination and month, with holiday and school-calendar effects handled explicitly, since those drive most of our variance. COVID years are in our history and will wreck a naive model. How you handle that period is something we will look at closely. Six years of booking data.

We contract about 400 farmers and commit to volumes before harvest. When we get it wrong we either buy expensively on the open market or leave farmers with unsold produce. We want per-plot yield estimates from weather history, soil test data, sowing dates and past yields, updated through the season as the picture changes. We would rather have an honest range than a confident point estimate. Our decisions are about how much risk to carry, so the uncertainty is the useful output. Five seasons of data across three crops.

We have four models in production, all deployed differently, none monitored. Two were trained by someone who has left and we cannot reproduce them. We want a standard path: experiment tracking, a model registry, containerised serving, and drift and performance monitoring that alerts. Retraining should be a documented pipeline, not an afternoon of someone's memory. This is platform work rather than modelling. Please send infrastructure you have built, not notebooks.

We know D1 retention and nothing beyond it. We do not know where in the first session players stop, or which early choices correlate with sticking around. We want event data modelled properly, then funnel and cohort analysis: where the drop-offs are, and whether the tutorial is the problem we suspect it is. We have raw events in BigQuery, unmodelled and somewhat inconsistent — cleaning them is part of this. Output should be a dashboard plus a written read of what it says.

Last-click gives all the credit to brand search, which is obviously wrong and yet drives our budget decisions. We want a data-driven attribution model over our touchpoint data — Markov or Shapley, argue for one — and a comparison against last-click so the difference is visible and arguable. We are not expecting truth. We are expecting a defensible second opinion that changes how we allocate. About 18 months of touchpoint data, roughly 40 million rows.

We run experiments and call them when the numbers look good, which means we are shipping noise. We want a proper framework: sample size calculation before launch, fixed horizons or sequential testing done correctly, guardrail metrics, and automated readouts that state significance honestly. We would like the tooling to make peeking hard rather than merely discouraged, because discouragement has not worked. Please include how you would handle multiple comparisons and low-traffic surfaces.

We deliver from six outlets to about 900 addresses a day using routes a supervisor plans each evening. We want routing that accounts for vehicle capacity, delivery windows, and the fact that perishables have a hard time limit in an unrefrigerated vehicle. Realistic travel times matter — straight-line distance produces routes that look efficient and fail in Pune traffic. Deliverable is a service our dispatch tool can call, plus a way for the supervisor to override and re-solve.

Music in one studio bleeds into the class next door, and both rooms have enough reverb that trainers are shouting by the third class. We want a treatment plan per room — absorption, placement, materials — with a specification a contractor can quote against and an expectation of what it will achieve. We also have a neighbour complaint about low-frequency transmission through the floor, which we suspect is the harder problem. Site visits in Pune required; drawings available.

We oversell about 30 orders a week because availability is read from a table updated by a batch job. We want a service that reflects reservations, in-flight carts and returns in near real time across web, app and two marketplaces. The hard part is concurrency: two customers taking the last unit within the same second, and a marketplace that reads stock on its own schedule. Please address that specifically. Overselling is worse than underselling for us — err conservative.

We hire in small batches and onboarding is whoever has time that week. New joiners take about three months to be useful and tell us the first fortnight is confusing. We want a structured 30-day programme: what they learn, in what order, who owns each part, and how we check it landed. Mostly remote team across three cities, so it cannot depend on someone sitting next to them. Deliverable is the programme, the materials, and a manager-facing checklist.

About 4% of our players ever buy anything, and we currently show the same store prompt to all of them at the same moment. We want a model that estimates purchase propensity from the first two sessions, so we can stop interrupting the 96% and time the offer sensibly for the rest. We are not trying to squeeze whales. Showing fewer prompts to people who will never buy is most of the value here, and we would rather the model be used to remove friction than add it. Event data in BigQuery, roughly 240,000 players.