Launching local LLM llama.cpp + analysis of attention and activations (Windows)

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.

I'm looking for a performer to build an AI-agent system that automates the marketing and sales pipeline: from content generation to lead segmentation in CRM and hypothesis analytics. Below are the tasks grouped by functional purpose. How many services/agents will be in the final architecture and on which stack is up to you, based on your own experience. The main thing is that the solution covers all the tasks below, is functional, maintainable, scalable, and must have a convenient mechanism for review/approval of results by a person before publication or launching ads. Block 1. Content Generation (Multichannel Copywriting)One or several generating modules that produce texts for various formats based on project materials (course programs, interviews, broadcasts): Landing page texts Email sequences: warming chains, newsletters Posts in Telegram bot/channel Advertising texts for Meta, including variants for different hypotheses Content plan and posts for Facebook / Instagram / Telegram Blog articles (including based on transcribed videos — see Block 3)Block 2. Production of Final MaterialsTransforming the finished text into a final artifact ready for publication: Lead magnets — layout and assembly of a ready PDF (checklists, guides) with Canva/Figma integration for editing Meta advertising creatives — static visuals with resizes for campaign formats Landing pages on Framer — page structure, CMS filling, assembly of a ready page for launch.Block 3. Video → Blog Transcription of videos (broadcasts, interviews, workshops) Cutting the transcript into articles that lead into the funnel Publication in the blog on the main site (there is currently no blog — possibly a blog section needed on Framer CMS; open question: can the system create it itself, or is this a separate task)Block 4. Integration with Meta Advertising Cabinet Uploading finished texts and creatives to the cabinet Tagging ads/campaigns by hypothesesBlock 5. CRM and Lead Routing Automatic segmentation of leads in the Telegram bot (attended/did not attend the event, funnel branch, offer) Transferring segments to CRM Auto-tagging new leads in CRM when entering the funnel (through registration)Block 6. Hypothesis Analytics Formation of a table/dashboard of hypotheses: costs/results for each funnel, lead magnet, creative. Periodic AI analysis with recommendations: what to scale, what to stop Evaluation of landing page conversion ratesWhat We Expect in the Response Architecture Vision — how would you break these 6 blocks into agents/services, on which stack (orchestration, generation, data storage). Portfolio/Cases — examples of similar automation systems (marketing, leads, content generation). Estimated assessment of timelines and costs, preferably step by step. Questions about the Terms of Reference, if something is unclear.

We are looking for an experienced Python/AI developer (or a small team) to create an AI service based on a Telegram bot. The project automates the acceptance of orders from masters and service centers on an active B2B marketplace for spare parts for mobile devices and electronics (catalog of ~30,000 SKUs). What the bot should do:Multimodal input: accept text, voice messages (Whisper API), and photos of parts/markings (GPT-4o Vision), recognize nomenclature, revisions, and part numbers.Smart search: match masters' jargon (battery, accumulator, display, original) with official names through vector search (Vector DB).Interactive UI: when there are several quality options (Original, High Copy, etc.) — group them into one message, display inline checkboxes with prices and stock, recalculate the cart "on the fly".Cascading clarifications (Slot Filling): when faced with an ambiguous request, clarify parameters step by step (model → color → revision).Order processing: automatically send the completed JSON order package to the website via API after confirming the cart.Stack: Python (FastAPI), Aiogram 3.x, OpenAI API (GPT-4o + Whisper), Qdrant, Docker + docker-compose. Implementation period: 5–7 weeks.⚠️ Mandatory condition for response: To filter out auto-responses from spam bots, please indicate in your message: which embedding model you will choose for vector semantic search across 30,000 technical SKUs and why? Responses with template text that do not answer this question will be automatically rejected.

We are looking for a specialist or a team to implement AI solutions in the communication of the CT and MRI medical center. Requirements: integrate AI into telephony for handling incoming and cold calls; ensure natural Ukrainian-speaking communication with minimal pauses; integrate an AI assistant into the Binotel chat or propose an effective alternative; automate responses to inquiries, initial consultations, and scheduling for examinations. Important: the medical field requires accurate information gathering before CT or MRI, particularly regarding the examination area, preparation, referrals, and possible contraindications. The solution must be empathetic, professional, and allow for the transfer of complex cases to an operator. In your response, please send relevant cases, a description of the proposed solution, and estimated timelines. Experience in medical projects will be an advantage.

It is necessary to develop an application in MATLAB that can process images/videos, identify individual objects, analyze their characteristics, and, if necessary, use machine learning methods to automate the analysis. Desirable: confident knowledge of MATLAB; experience in Computer Vision / Image Processing; experience with Machine Learning;