Find Work

Our catalogue has 12,000 products and the attribute data is a mess — colour, sleeve length, neckline and pattern are either blank or wrong on roughly a third of them. Filters on the site are useless as a result. We need a model that reads the product photo and predicts those four attributes, plus a confidence score so we can route the uncertain ones to a human. We have about 3,000 correctly tagged items to train on. If that is not enough, say so up front rather than after you start. Deliverable is the model, the inference code, and a short handover doc our backend team can work from.

We have around 400 pages of product documentation in Confluence and a support team that answers the same twenty questions every week. What we want: a retrieval chatbot our support agents can query internally. It should cite the source page for every answer and say "I don't know" rather than guess — a confidently wrong answer is worse than no answer for us. Scope: - Ingest and chunk the Confluence export (we will provide it) - Embeddings + vector store, your choice, but justify it - A simple web UI, no design work needed — we have a design system - Answer must link back to the source page We are not looking for a wrapper around ChatGPT with a system prompt. Please tell us how you would handle chunking and evaluate retrieval quality.

We run six retail outlets selling fresh produce. We over-order and throw away roughly 12% of stock, or we run out by 6pm. Both are expensive. We have two years of daily sales by SKU by store, plus festival dates and weather if that helps. We want a weekly order recommendation per store, and — more importantly — a plain explanation of what drives it, because our store managers have to trust it enough to actually follow it. A black box they ignore is worth nothing to us.

We are opening our fourth studio and our branding still looks like the one my co-founder made in Canva in 2022. We need the full set: primary and secondary logo, colour system, type system, and a guidelines document a printer and a web developer can both work from without calling us. Our audience is 25-40, mostly women, in Pune and Mumbai. We are not a heavy-lifting gym and we do not want to look like one. Think calm and precise rather than aggressive. Please show brand work, not just logos. We want to see how you think about a system.

We process a few hundred vendor invoices and bank statements a week and extract line items by hand. We want that down to a review step rather than a data entry step. The documents are Indian formats — GST invoices, bank statements from six or seven banks, all PDF, some scanned. Layouts vary a lot. What matters most to us is that the model is honest about uncertainty. We would rather it flags 20% for review than silently gets 5% wrong, because a wrong figure in a reconciliation costs us far more than a manual check. Please share something comparable you have built. Accuracy claims without a test set do not help us judge.

We get 60-80 inbound calls a day and most are not a fit. We want a voice agent that answers, asks four qualifying questions, and either books a slot on our calendar or politely closes the call. Must handle Indian English and Hindi. It has to sound like a normal call, not an IVR tree — if it cannot understand something twice it should hand off to a human rather than loop. Hourly because we expect to iterate on the script a lot in the first month. Please quote an honest hourly rate and a rough range of hours.

About 90,000 lines of Node that started as an MVP in 2021. It works and it makes money, and it is now slow to change — a small feature touches six files that should not know about each other. We are not looking for a rewrite. We want incremental extraction of clear domains behind interfaces, with tests going in ahead of each move, shipping continuously the whole time. This suits someone who has done this before and knows where it goes wrong. If your instinct is to start with a new repo, we are probably not a fit. Long engagement, hourly, roughly 20 hours a week.

We split payments to about 200 partners using Razorpay Route and reconcile settlements in a spreadsheet. It has been wrong twice, and both times took a day to find. We want a service that ingests settlement reports, matches them against our transfers, and shows exactly what is unmatched and why. Idempotent webhook handling, and an audit trail that survives a replay. Money code, so we care more about correctness than speed of delivery. Tell us how you handle a webhook arriving twice, and a settlement that arrives before the transfer record.

Three playable characters, 32x32 pixel art, each needing idle, run, jump, fall, attack and hit states. Style reference is Celeste and Dead Cells — readable silhouettes, limited palette, animation that reads clearly at speed rather than at rest. We will share our existing tileset so the characters sit in the world properly. Deliverables as Aseprite files plus exported sheets with a JSON atlas. If you have shipped sprite work in an actual game, lead with that.

Nine SKUs across three flavour families, launching in modern trade and quick commerce. The design has to work at two very different sizes: a 60g pack on a shelf, and a 200x200px thumbnail on Blinkit where the flavour has to be readable instantly. Most designs we have been shown fail the second test. We need print-ready artwork on our dielines, with FSSAI panel and nutrition table laid out correctly. Experience with Indian food packaging compliance is important — we do not want to discover a labelling problem at the printer.

Our Shopify theme is slow — 4.5s LCP on 4G, which in India is most of our traffic — and we have run out of things to fix inside the theme. We want a headless storefront on Next.js against the Storefront API. Catalogue, PDP, cart, checkout handoff, and search. The number we care about is LCP under 2.5s on a mid-range Android over 4G, measured on a real device and not on your laptop. Please tell us how you would get there before you tell us what you would build it with.

We post four to six times a week and every post looks like a different company made it, because a different person does make it each time. We want a template set in both Figma and Canva — carousel, single post, story, and reel cover — that our intern can use without breaking the brand. The constraint that matters: it has to survive being edited by someone with no design training. If it falls apart the moment the caption is two lines longer, it has not solved our problem.

Our dashboard has grown feature by feature for three years and it shows. Users tell us it is powerful and confusing, which we agree with. This is a redesign of the core analytics screens — roughly 14 screens — plus a component library our front-end team can build against. Not a fresh coat of paint: we expect you to question the information hierarchy. We will give you access to session recordings and our support tickets. The problems are well documented; we are not asking you to guess. Please include a SaaS product in your portfolio. Marketing site work does not tell us what we need to know here.

We have a working web product with about 8,000 monthly users asking for an app. The API exists and is documented. Scope is the core flows only — auth, browse, the main create flow, notifications, profile. Not everything the web has. We would rather ship six screens that feel right than twenty that feel like a wrapped website. Both stores, and we need you to take it through review rather than hand us a build and disappear. Budget assumes you handle submission. Please link apps that are live, with their store listings.

New studio opening in Kothrud, Pune in about six weeks. We need trials booked, not impressions. Current state: an Instagram page with 4,000 followers and no paid experience at all. Pixel is not installed. Assume nothing is set up, because nothing is. We want someone who will tell us when the offer is the problem rather than the targeting. Our last agency ran ads for four months and reported reach. Monthly ad budget is ₹60,000-80,000, separate from your fee.

12,000 SKUs, a lot of near-duplicate category pages, and organic traffic that has been flat for a year while the catalogue tripled. We want a real technical audit — crawl budget, faceted navigation, canonicals, schema, internal linking — and a prioritised roadmap our developers can actually execute, with effort and expected impact against each item. What we do not want is a 60-page automated report. We can run those ourselves. We are paying for the judgement about what to do first.

About 600 published articles, decent organic traffic, and a WordPress install that falls over whenever a post does well. We want a headless setup — editors stay in WordPress, the front end moves to Next.js. The part we are nervous about is SEO. Every URL must either stay identical or 301 correctly, and we want to see the redirect map before you cut over. We have lost traffic to a botched migration before and would rather move slowly.

Stock counts in Zoho Books and in our POS drift apart daily and someone reconciles them by hand every evening. We want a sync service — near real time, both directions where it makes sense, with a clear rule for which side wins on a conflict and a log we can read when we do not believe it. The hard part is not the API calls. It is what happens when the POS is offline for two hours during a sale rush. Please tell us how you would handle that.

Our docs were written by the engineers who built the features, which means they are accurate and unreadable to the people who buy the product. About 60 pages. We want them restructured around what a user is trying to do rather than around our menu structure, and rewritten so an operations manager can follow them without asking support. Accuracy still matters — you will have access to the engineers, and we would rather you ask than guess. Please send documentation you have written, not blog posts.

Our founder has strong opinions about SME lending in India and no time to write them down. The arrangement: a 45-minute call every fortnight where she talks, and you turn that into three or four posts a week in her voice. Not thought-leadership filler — she would rather post nothing than post something generic. You need enough finance literacy to catch it when a transcript is loose about a number. That is the part most applicants fail. Ongoing, roughly 6-8 hours a week.