Разработка ML модели категоризации текста, картинок

Task: one dashboard with all business metrics — advertising, funnel, payments, manager performance, revenue planning. Data is pulled automatically via API. Scope: only the YCL direction (employment in Europe). Kommo has other directions — only YCL funnel deals will be included in the repository (filter by funnel/tag to be agreed upon).1. Data Sources (Integrations) Kommo CRM — leads, deals, funnel stages, responsible persons, sources, dates of transitions between stages (must keep history), reasons for refusals, custom deal fields (see point 2). Stripe — payments, amounts, statuses (success/failure/refund), linked to deals. Meta Ads — expenses, impressions, clicks, CPL, leads by campaigns (currently operational). Google Ads, Reddit Ads, LinkedIn Ads — planned; architecture — extensible connectors without core rework. SEO/organic— Google Search Console + GA4. Cross-link: traffic source → lead in Kommo → payment in Stripe (UTM, deal ID in Stripe metadata — propose the mechanism). 2. Mandatory Cuts (Deal Fields in Kommo) Each metric must be filtered/grouped by: Client Citizenship (Kenya, Nigeria, India, etc.). Residence Status: lives in their country / expat (already in Europe). These are two different segments with different cycles, conversion rates, and checks. Country of Placement / Service: Poland, Serbia, Slovakia, Germany (ZAV). Manager, team, traffic channel, period. If any fields are missing in Kommo — the executor indicates which fields need to be added, the client adds them.3. Funnel and Leading Indicators Data by funnel, for each stage — summary and leading metrics: Traffic → lead: leads, CPL by channels + day-to-day expense/click dynamics. Lead → qualification: conversion + first response speed, touches/calls to the manager per day, unanswered leads. Qualification → contract/invoice: conversion + sent offers, stalled deals (days in stage above norm). Invoice → payment: payments, average check + unpaid invoices, failed payments. Summary: revenue, ROMI by channels, run rate to monthly plan. 4. Deal Cycle Average and median lead → payment cycle (business benchmark ~4 weeks), cycle trend over time. Breakdown of cycle by stages (how many days a deal sits at each stage) — to see which stage is dragging. List of deals that have stalled at a stage longer than normal. Cycle breakdown by segments: citizenship, residence status, country of placement, manager. 5. Early Warning of Decline (Key Block) Since the cycle is ~4 weeks, today's leads = payments in a month. The system must: Compare leads/qualifications of the current week with the moving average (4 weeks) and issue an alert if there is a downward deviation: “leads -X%, with a 4-week cycle expect a payment decline in the week [date].” Build payment forecast for 4 weeks ahead from the current pipeline: deals at each stage × historical conversion of the stage × remaining cycle. Highlight in red weeks where the forecast is below plan — with time to react. 6. Additional Payments and Sales Planning In the Kommo deal card, the date and amount of the planned additional payment are stored. The system must: Collect a calendar of upcoming additional payments: total expected, by weeks/months. Highlight overdue additional payments (date passed, no payments in Stripe) — a separate list for follow-up. Calculate the monthly plan as: plan − already paid − scheduled additional payments = how many new sales are needed (in money and in deal units at average check). Weekly schedule: additional payments + forecast of new payments against the weekly plan. 7. Manager Performance Daily snapshot for each manager: touches/calls, conversations, sent offers, payments — for each day separately, with a chart over the period. Progress on personal plan compared to monthly pace (ahead / on pace / behind). Benchmarking with colleagues. 8. Visualization and Roles “Traffic lights” (green/yellow/red) for key metrics relative to norms/plans; progress scales; trend graphs; mobile adaptive. Roles: CEO — everything; COO — entire funnel and managers; team lead — their team; manager — their metrics and position relative to colleagues. 9. Reports and AI Automated reports on schedule (daily summary, weekly report) in the dashboard and/or messenger. Free-form queries (“how has CPL from Meta changed over 2 weeks?”) — LLM over the repository. Alerts in the red zone and according to the rules from points 5–6. 10. Technical Expectations and Staging Repository (PostgreSQL/BigQuery or equivalent) + ETL: Kommo webhooks + periodic synchronization (15–60 min). Frontend: custom or BI tool — propose with justification; requirements for roles, traffic lights, forecasts, and AI queries must be implementable. Stages: (1) audit and metrics map → (2) MVP: Kommo + Stripe + Meta, funnel, traffic lights, roles → (3) deal cycle, early warning, additional payments and plan → (4) SEO, AI reports, alerts → (5) new advertising channels. Payment is staged, with a demo for each stage. In the response, indicate: similar projects (end-to-end analytics), stack with justification, timeline and cost estimates by stages, monthly ownership cost (hosting, tokens, licenses).

Task: deploy an LLM service that knows all the company's documentation and answers questions from the sales department managers. Current situation: the client has independently assembled a prototype (a separate project with uploaded company information, hosted on a server), but the information from the database is not transmitted to the model — likely, there is an issue with the API. We will provide the code and access. The first step is an audit: fix the existing setup or justifiably rebuild from scratch. Required functionality: Upload all company documentation: description of each service, regulations, FAQ, pricing (all materials will be provided). Answers strictly based on the uploaded documents (RAG). The model does not invent facts; if the answer is not in the database — it honestly informs about it. Access for managers via a link (web interface), with authorization. Scenarios: the manager asks any question about the company's work; inserts the client's question "as is" and receives a ready answer for sending; finds the necessary regulation/report by request. Knowledge base updates without a developer (uploading files through the interface or a connected folder). English language. History of requests for quality control. Technical expectations: LLM via API (Claude/OpenAI — propose with a cost calculation for tokens), RAG pipeline (vector database, embeddings), hosting on our server or in the cloud, HTTPS. The architecture should allow for future connection of the assistant to the analytical data warehouse (parallel project). In the response, indicate: examples of similar RAG projects, stack, timeline, cost of work, and estimated monthly ownership cost (tokens + hosting).

Creative Marketer / Ad Campaign Creator for Merivy — an AI-powered platform for beauty & aesthetics businesses (with a mascot!) Who we are We're a small startup building Merivy — booking and client management software for aesthetic clinics, beauty salons, barbershops, and other appointment-based businesses. At the heart of the product lives Merv — our AI agent (and green hand-shaped mascot ) who helps owners run their business: he sets up bookings, manages services and schedules, answers questions, celebrates wins, and generally feels like a team member, not a chatbot. What we're looking for A creative person who can turn this into a campaign people actually remember. Our reference for energy and tone is the media presence of viktor ( meet viktor) — we're a very different product, but we love how they talk to their audience: bold, human, funny, zero corporate blah-blah. We don't want a copy. We want that level of craft, with our own voice. The message we need to land Merivy helps you manage your clients, keep them happy — and most importantly, keep them coming back. Merv is the face of that promise: the little green teammate who never forgets a client, a booking, or a birthday. What you'll create A campaign concept built around Merv as the brand character (his voice, personality, running jokes) Scripts / storyboards for short-form video ads (IG Reels, TikTok) aimed at salon & clinic owners Static ad creatives and hooks for paid social Messaging we can reuse on the landing page and in the product You're a great fit if You've built campaigns or content for SaaS, beauty, or local-business audiences You can show us one thing you made that a stranger would send to a friend You think in characters and stories, not just "features and benefits" To apply Send 2–3 examples of your work and one sentence: how would Merv introduce himself to a salon owner in an Instagram ad? That one sentence matters more than your CV.

We are looking for a specialist who can develop and implement AI agents for sales automation and build a complete customer acquisition funnel.Tasks Develop an AI agent based on ChatGPT (or similar LLM). Set up a Telegram bot with AI. Integrate the bot with CRM. Build an automated sales funnel. Set up lead collection from Instagram, Facebook, TikTok, and the website. Develop communication scripts with users. Create quizzes and tests for audience segmentation. Set up personalized recommendation delivery. Organize automatic appointment scheduling through a calendar. Set up automated email and Telegram sequences. Integrate payment systems (if necessary). Prepare analytics on conversion at each stage of the funnel.Experience with the following is preferred ChatGPT API / OpenAI n8n Make (Integromat) Zapier Telegram Bot API CRM (HubSpot, GoHighLevel, Bitrix24, AmoCRM, etc.) Meta API WhatsApp Business API Calendly StripeWhat we want to achieve A ready-made system that: automatically communicates with potential clients; identifies their requests and needs; segments by interests; offers the appropriate product; books consultations or sells products; transfers data to CRM; requires minimal human involvement. When responding, please send: examples of implemented AI agents; examples of automated funnels; a list of technologies used; cost and timeline for project implementation.

We are looking for a highly skilled AI application engineer and full-stack backend developer to build a production-ready AI-powered document validation, refinement, and approval workflow. This is not a simple prompt engineering role. We need someone who can design and implement a real AI application with strong backend architecture, Claude API integration, structured validation logic, audit trails, secure data handling, and human-in-the-loop review workflows. The system will act as an intelligent quality assurance layer for submitted reports and documents. It should review completed submissions, identify issues, improve content quality, apply business rules, protect sensitive information, and either approve the document automatically or route it for human review. The developer will be responsible for building a workflow that can: Pull completed documents, reports, or submissions from an external platform via API Analyze the full document, including structured answers, ratings, selections, narratives, comments, and free-text fields Perform semantic audits to detect logical conflicts, contradictions, missing information, vague statements, unsupported claims, or incomplete sections Validate that structured responses and written content are consistent with each other Apply custom validation rules, editorial guidelines, formatting standards, tone requirements, and business logic Detect, tokenize, mask, or securely handle PII, confidential data, and sensitive security-related information before AI processing where required Rewrite and enhance narratives, comments, and document sections for grammar, clarity, professionalism, consistency, and readability Preserve the original meaning, observations, and intent while improving the final output Standardize writing style across documents without making every report sound generic or over-normalized Flag content that appears inconsistent, fabricated, vague, incomplete, sensitive, or requiring human review Generate specific validation notes explaining why a document failed review and what needs to be corrected Automatically generate clarification or revision requests when more information is needed Support approval workflows where documents are: Automatically approved when confidence thresholds are met Routed to a human editor or validator for review Returned to the original submitter for revision or clarification Maintain a complete audit trail showing: Original submission Tokenized or masked sensitive data events AI findings and recommendations AI-rewritten content Human edits Approval or rejection decisions Final approved version Write approved and validated content back to the source platform through API integration The role also requires building an editor and final-decision workflow. Human reviewers should be able to inspect the AI’s findings, compare original and revised content, make edits, approve changes, reject recommendations, and finalize the document before it is sent downstream. Ideal experience includes: Strong Claude API / Anthropic API integration experience Experience building AI-powered document review, validation, editing, or compliance workflows Strong backend architecture skills Full-stack development ability Experience with API integrations, webhooks, queues, job processing, and database design Ability to design structured AI outputs, confidence scoring, rule-based validation, and human-in-the-loop review Experience with PII detection, tokenization, masking, encryption, access control, and secure AI data handling Experience building secure audit trails and approval systems Strong understanding of prompt design, but also the engineering skills to turn prompts into a reliable production system We are looking for someone who has already built serious AI applications, not someone who only writes prompts. The right person should be able to design the architecture, integrate with external APIs, manage document processing logic, protect sensitive data, build the review interface, and deliver a reliable workflow that can be used in production.