Implementation of CRM for a distributor of glass structures

A small adjustment is needed in the program that generates print orders. Task: We have added to the assortment soft glass with print. In the orders, the name of the print will be highlighted with the symbols #...# for correct recognition. Example: Soft glass, #White abstraction# 2.0mm, Shape: Rectangle... Requirements: If the text of the order contains a fragment between the symbols #...#, the program should recognize it as an order with print; when generating a PDF or print sheet, such an order needs to be highlighted in a different color (any noticeable color, to be agreed upon during execution); the rest of the program's logic should remain unchanged. An example of the order is attached below for testing.

About the project: We are launching a B2B service for end-to-end analytics and advertising campaign management for targeters and media buyers. The product will work with the official Meta API. The main technical challenge and focus of the project is the virtuoso handling of Facebook limits, traffic routing between our application pool, and a robust infrastructure protection system against the blocking of gray advertising accounts. We already have a detailed technical specification, the database architecture is described, the load balancer logic, and interface requirements. We are looking for a contractor who will take this for turnkey implementation (backend + frontend dashboards). What needs to be done (Key tasks): Integration with Meta API: Set up user authorization and regular asynchronous parsing of advertising account statistics. Requests should be sent exclusively in batches to save limits. Auto-unlinking system for accounts: Write a module that continuously monitors the statuses of advertising accounts. If an account gets banned, the system should automatically revoke the access token within 60 seconds to protect our application from Meta sanctions. Proxy infrastructure: Implement tunneling of all API requests through a SOCKS5 pool. A strict binding of a specific user token to a static IP address is mandatory. Smart routing: Create an algorithm that will distribute bindable advertising accounts among several of our Facebook applications in specified proportions to reduce risks. Interface development: Create a client dashboard with a summary statistics table and an advanced admin panel for manual management of user limits, application bindings, and proxy pool. Expected technology stack: Backend: Python, FastAPI. Asynchronous tasks: Celery, Redis. Databases: PostgreSQL (or ClickHouse for statistics, at your discretion). Frontend: Vue.js or React (you can use ready-made UI libraries and dashboard templates, focusing on functionality rather than complex design). Requirements for the contractor: Confident experience with Meta Graph API and Marketing API. You should understand how sliding limits work, how to read load headers, and how to work with tokens. Understanding the specifics of traffic arbitration. The words "billing," "account ban," "business manager," and "farm" should not raise questions for you. Experience in building asynchronous parsers and working with proxy servers at the network request level. Willingness to work according to a clear technical specification and deliver the project in stages. Conditions: Collaboration format: Project work Budget: Discussed individually based on your assessment of the technical specification. Payment: Staged, tied to checkpoints. How to respond: In your cover letter, be sure to indicate your experience with Meta API, attach links to similar projects (or describe their functionality if they are under NDA), and provide an estimated price range and timeline for developing such a system from scratch. Responses without a description of relevant experience with Facebook API will not be considered.

We are looking for a specialist in LLM, RAG, and prompt engineering for auditing and improving an already created AI assistant for contact center operators of a network of medical centers. This is not a development from scratch. Currently, the assistant operates in the ChatGPT environment and uses: its own skill with instructions SKILL.md; a knowledge base in Project Sources; structured Markdown files automatically generated from CSV exports of the medical information system; separate indexes of prices, performers, departments, packages, equipment, and recommended service combinations. The database contains approximately: 2,700+ medical services; 90+ packages and complexes; 300+ recommended combinations; 600+ surgical interventions; prices by various departments; performers, addresses, preparation, equipment, and other reference information.What the assistant should do Upon the operator's request, the assistant should quickly provide a verified response: whether the required service is provided; the exact code, name, and price; in which departments it is available; which doctors or other specialists perform it; how to prepare; which package or complex is more advantageous to offer; which accompanying services are advisable to suggest; the sequence of comprehensive patient registration; what cheaper or alternative options are available; for operations — separately the base price and the estimated total cost of the surgical treatment case. The assistant should not invent prices, performers, preparation, medical indications, or transfer information between similar services.Existing problems The system is already operational but requires increased search stability and response quality. In particular: the model sometimes finds the main service but misses recommended combinations; does not always extract individual fields from large Markdown files; can find the base price of an operation but miss the total cost of the surgical case; results depend on the structure and size of the files in Project Sources; indexes, source routing, and search rules need optimization; it is necessary to ensure equally high-quality responses to short, inaccurate, and conversational operator queries. For example, a simple query "cholecystectomy" should immediately return available options for the operation, codes, base prices, total treatment costs, departments, performers, and related services.Specialist tasks Conduct an audit of the current SKILL.md, the structure of the knowledge base, and the search logic. Analyze the reasons for data omission during retrieval. Propose an optimal knowledge base architecture for ChatGPT. Improve or rewrite SKILL.md. Optimize the structure of Markdown files and compact indexes. Set up mandatory searches: packages and complexes; recommended combinations; prices by departments; performers; the cost of the surgical treatment case. Check the knowledge base generator from CSV exports and improve Python scripts if necessary. Create a set of control queries and criteria for evaluating responses. Conduct testing on real contact center scenarios. Provide final documentation for further updates and system support.Expected result We expect to have a stable assistant that: responds in Ukrainian; does not miss critically important data; returns only information confirmed by the knowledge base; correctly distinguishes between services, packages, and recommended combinations; shows code, name, price, department, and performer; for operations separates the base price and total treatment cost; offers the operator a specific scenario for further patient registration; works stably after subsequent updates of CSV exports; operates relatively quickly.Requirements for the performer A specialist with practical experience is needed: ChatGPT Projects, Custom GPT, or ChatGPT Skills; LLM, RAG, retrieval, and semantic search; prompt engineering; designing knowledge bases for language models; Markdown, CSV, JSON/JSONL; Python for data processing and transformation; testing the quality of LLM responses. Experience with medical information systems, contact centers, or large service catalogs will be an advantage. We are looking for not just a prompt author, but a specialist who understands the limitations of searching in large sources, context fragmentation, and ways to build reliable indexes.What to provide in the proposal Please briefly indicate: Your experience with ChatGPT, RAG, or corporate knowledge bases. Examples of similar implemented projects. How you would approach diagnosing the omission of individual fields in large files. Estimated timelines and costs for the audit and refinement. Whether you are willing to sign a confidentiality agreement. Personal data of patients will not be transferred within this project. The final cost of the work will be agreed upon after clarifying the Technical Task between the Customer and the Performer.

Brief I need a program in Python that runs on my PC (Windows) and creates faceless videos in the format of "voiceover + changing visuals from photos and clips" — like historical documentaries on YouTube (I will attach an example separately). I input a topic → the program writes a script, voices it, selects video/photos from free archives for each piece of text, edits them together → outputs a ready MP4. For me only. No website, no users, no sales. One user — me. How it works (step by step) 1. Input. A simple window opens. I enter the topic of the video and choose a voice from the list (**the list of voices is automatically pulled from ElevenLabs via API** — available on my account). I click "Create." (Second mode: insert a ready script instead of generating.) 2. Script. The program writes a script on the topic of the specified length via LLM API (OpenAI/Anthropic, key in settings). 3. Scene breakdown. LLM divides the script into scenes and returns for each: scene text; visual type: video or photo; search query (a detailed phrase of what should be in the frame); highlight mark + importance rating 1-10 (for intro, see below). 4. Voiceover. The text is sent to ElevenLabs API with the chosen voice → audio. The video sequence is cut to match the length of the audio for each scene. 5. Video/photo selection — from free sources via API (see the list below). A specific request is formed for each source (for accuracy). If one doesn't yield results — it tries the next. 6. Verification (maximum 3 steps per scene). Step 1: the program takes the first found option (video/photo) based on the request. Step 2: sends the frame to LLM — "does it fit the scene?". If it fits → stop. Step 3 (if it doesn't fit): the program switches to searching for photos (it's easier to find photos with an exact request than videos — this guarantees relevance) and takes it with enhanced motion (zoom + pan). No more than 3 steps per scene — this saves the LLM budget and ensures that the frame is relevant. 7. Assembly via ffmpeg/moviepy: clips and photos timed to the voiceover, photos are animated with zoom (Ken Burns effect), voice on top, simple transitions. Output: MP4 1920×1080. Video sequence rules (important — the program is responsible for this) First 60 seconds — intro teaser: a montage of the most impactful clips from the entire video (scenes with the highest importance rating that contain video) under a separate introductory text from LLM ("in this video you will learn..."), frames without explanations, creating intrigue. Then a transition to the main part. Alternation: video insert at least every ~6 seconds, no many photos in a row. Video share: at least ~40% of the time — live clips, the rest — photos with zoom. Frame length: 4-6 seconds (both photos and videos). No rapid cuts, no prolonged static shots. For purely historical topics where there is no video — photos with enhanced motion (zoom + pan). Sources (all free, with API) Modern video + photos: Pexels, Pixabay. Historical/archive (public domain): Wikimedia Commons, Archive.org, Library of Congress, Europeana, NASA, Smithsonian Open Access, Flickr Commons, openverse. Each source is a separate module, easy to add new ones. Use only public domain / free licenses with the right for commercial use. No parsing of other YouTube/sites, movie clips, images "from Google".Uniqueness of selection To ensure videos do not match others: take a random clip from the top results (not the first), maintain a database of already used clips (do not repeat), optionally — light processing of the clip (crop/mirror/speed). Options (enable/disable in settings) Photos only — if enabled, the video is assembled PURELY from photos, without video clips. Each photo MUST have motion (zoom and/or pan, Ken Burns effect) — even in this mode, there should be no static "dead" frames, minimal dynamics always. If disabled — standard mode (photos + video clips alternating, as per video sequence rules). No voiceover — if enabled, the video is assembled based on the text WITHOUT generating voice: the code DOES NOT call ElevenLabs and does not overlay voice (the video sequence is selected based on the text of the scenes, timing of frames — according to rules/parameters, without reference to audio). If disabled — it automatically generates voiceover based on the text through ElevenLabs, as usual. Atmospheric overlay — if enabled, a semi-transparent layer with floating particles/dust/glowing bokeh/light fog (particle/dust/bokeh/fog overlay, screen/add mode) is applied over the entire video sequence to make the frames look alive and cinematic. When installing the program, a set of 5-8 popular overlays (particles, dust,copybokeh, fog, light cinematic "grain") is placed in a local folder — I choose the needed one from the list. Adjustable transparency/brightness of the overlay (slider 0-100%), so the effect is neithercopytoo dull nor too pronounced — I adjust the strength myself. Ideally, the overlay should also be applicable to ALREADY finished videos separatelycopy(post-processing: take a ready MP4 → choose overlay → set transparency → save), not just during assembly. Where to get overlays for packaging (free license): Pexels, Pixabay (queriescopy"particle overlay", "bokeh overlay", "dust overlay", "light leaks", "film grain"), Mixkit, Videezy. The performer selects 5-8 pieces and places them in the program folder. Subtitles (embed or separate .srt). Clip processing for uniqueness. Resolution/format, video length, video share, search depth.Technical requirements Python. Modular structure (sources and LLM — through interchangeable modules, to easily replace or add). All API keys — in the settings file, not in the code. Simple window (GUI at the discretion of the performer — Tkinter/PyQt), launched by double-clicking. README with instructions, clear logs, comments in the code.What I provide API keys (ElevenLabs, LLM, where registration is needed — I will arrange). I will pay for any fees myself. Examples of video references (I will attach) and examples of topics for tests.Acceptance (ready if) I launch → window → I enter the topic, choose the voice → "Create" → I receive a ready MP4. The video sequence matches the meaning of the text, alternating video/photos, intro teaser 60 sec, voiceover on top. Works with at least 6 free sources, with fallback between them. Frame verification through LLM: max. 3 steps per scene (found → LLM checked → if not, photo with motion as a sure bet). Uniqueness: randomization + database of used. Only legal sources. There is a README, runs from scratch.Delivery of results All source code — in open view (all files), without obfuscation + compiled working version. I can run it myself from the sources according to the instructions (README: installation, keys, launch). The code must be clean, commented, and understandable, so **any other programmer can continue working** on it if needed (not tied to the author). **All everyday work — through the interface (buttons, fields, sliders, dropdown lists), WITHOUT the need to touch the code.** All settings (topic, voice, options, overlay, folders, length, formats) are changed in the program window, not by editing files. The code in hand — only as my property and insurance, not as a way to control the program. All rights to the code after payment — mine.Disk space management (important) The program should not fill up the disk. Implement: After assembling the video, all intermediate files (downloaded clips, temporary pieces, audio cuts) are automatically deleted — only the ready MP4 remains on the disk. Cache limit (parameter in settings, e.g., 5 GB): when exceeded, old downloaded files are automatically deleted (starting with the oldest). I set the folder for ready videos and for temporary files in the settings. Show how much space is occupied, and a button "clear cache" manually.Please specify in your response Examples of similar works (ffmpeg/moviepy, working with stock/archive APIs, ElevenLabs/LLM). Proposal for GUI. Does the solution use a database (which and why) — or are local files sufficient. Timeline and cost.

We are developing a real-time integration with the external service Trainer. We send structured snapshots of the state, receive recommendations, and display them in a pop-up bubble. The task is stable and complete data transmission for the correct operation of the Trainer. We are looking for a developer for the chain: data processing → HTTP communication → overlay. We need people with the following skills Python is good, Java basics, API HTTP/JSON-APIs The project is 90% ready but there are some inconsistencies