Cold Calls assistant application using ChatGPT, Whisper, VOIP and text

Technical task: Program for automatic video assembly (local, for personal use) Briefly Program in Python, runs on my PC (Windows). Creates faceless videos in the format of "voiceover + changing visuals from photos and clips" — like historical documentaries on YouTube (references will be attached separately). I input a topic → the program writes a script, voices it, selects video/photos for each piece of text from free archives, assembles → outputs a ready MP4. For me only. No website, no users, no sales. Role division (important to understand) Claude (Anthropic API) — "brain": writes the script, divides it into scenes, selects the best frame by meaning. Works only with text and images. Code (ffmpeg/moviepy) — "hands": searches and downloads files from archives, cuts, glues, overlays sound, zooms, transitions, exports MP4. Claude does NOT edit video files — that’s done by the code. How it works (step by step) 1. Input. Simple window: I enter the topic of the video, choose a voice from the list (the list of voices is pulled from ElevenLabs via API). I click "Create". The second mode — insert a ready script. 2. Script. Claude writes a script on the topic of the specified length. 3. Scene breakdown. Claude divides the script into meaningful scenes (length by meaning, not fixed) and returns for each: scene text; type of visual: video or photo; search query in two formats at once (Claude provides both in one response, no extra queries or costs): expanded phrase — for stocks (Pexels/Pixabay), e.g. "18th century royal portrait, dramatic turning point"; brief keywords/names — for archives (Wikimedia, LoC, Archive.org, Europeana), e.g. "Marie Antoinette portrait". The code takes the appropriate format for each source. Claude generates the query based on the MEANING of the piece and the context of the entire video — even for abstract pieces without a clear object. There are no empty pieces "without a query"; highlight: true/false + importance rating 1-10 (for intro teaser). Meaningful scene ≠ frame. Claude divides the text by MEANING (so that the query is meaningful). The code then cuts this into visual frames of 4-6 seconds, adjusting the timing to the voiceover: short scene (≤6 sec) → 1 frame; long scene (e.g. 12 sec) → 2-3 frames of 4-6 sec each, with each — DIFFERENT, but close in meaning visual (same topic, different angles/frames), so that the thought does not hang on one static picture. The 4-6 sec rule applies to the frame, not to the meaningful breakdown — they are at different levels and do not conflict. 4. Voiceover. Text → ElevenLabs API with the selected voice → audio. The video sequence is cut to the length of the audio of each scene. 5. Frame selection (key process, see section below). 6. Assembly via ffmpeg/moviepy: clips/photos according to the timing of the voiceover, photos with zoom (Ken Burns effect), simple transitions, voice on top. Output: MP4 1920×1080. Frame selection (how video/photo is chosen for the scene) Search across all sources at once by the scene query. For each source, the appropriate query format is taken (see breakdown above): for archives — brief terms/names, for stocks — expanded phrase. The goal is to hit the topic, not a random picture. The code selects the top-5 candidates from all results by scores (scorer, see rules below). For video the code extracts 2-3 still frames from different points of the clip (beginning/middle/end) — so that Claude evaluates the clip by frames. Claude selects the best from 5 candidates by meaning (is the needed content realistically on the frame, does it fit the mood of the scene). Limit of 3 steps per scene: found a suitable one — stop; if nothing fits after checking — the code takes a more general photo by a simplified query (it will definitely be found) with enhanced movement. ALL frames are checked (the price difference is negligible, and quality is more important). Candidate selection rules (scorer) — to take good, not junk Relevance to the query (position in the source output) — main weight. Not used before — checked against a local database of used ones, do not repeat clips in my videos. Format — horizontal 16:9 prioritized. Quality — DIFFERENT for the type of source: modern stocks (Pexels/Pixabay) — filter out low quality, require HD; historical archives — low threshold/off, DO NOT lose film and documentary frames, even low resolution (their value is in the material itself, not in clarity). Clip length (for video) — is it enough to cover the scene. Diversity — do not take many identical frames in a row from one source, mix sources (so that the video sequence is diverse, not from one archive). Uniqueness (fewer matches with others' videos) Randomization of output (random from top-N, not always the first), search depth, local database of used ones (do not repeat clips in my videos), optionally — light processing of the clip (crop/mirror/speed). When unused frames for the query run out: the code first expands the search (other sources, deeper into the output, rephrasing the query). If that is exhausted — allows repeating previously used ones by priority: first those taken long ago (not in the latest videos), with mandatory processing for visual distinction. The video should not stall due to lack of material. Sources (all free, with API) Modern video + photo: Pexels, Pixabay. Historical / archival (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 a new one. The priority of sources depends on the topic (historical → archives first; modern → stocks first), configured in the settings. Only public domain / free licenses with the right to commercial use. No parsing of others' YouTube/sites, film clips, images "from Google". Video sequence rules Intro teaser (first 60 sec): a montage of the most striking frames from the entire video (scenes with the highest importance rating, where there is video) under a separate text-introduction from Claude ("in this video you will learn..."), frames without explanations — as intrigue. Then transition and main part. Alternation: video insert at least every ~6 sec, cannot have many photos in a row. Video share: at least ~40% of the time — live clips, the rest — photos with movement. Frame length: 4-6 sec. Do not flicker, do not hold static for long. For purely historical topics, where there is no video — photos with enhanced movement (zoom + panorama). Options (on/off in settings) Only photos — video purely from photos, without clips. Each photo must have movement (zoom/pan) — no static "dead" frames even in this mode. No voiceover — video is assembled by text without generating voice (the code does not call ElevenLabs). Otherwise — voiceover is done automatically. Atmospheric overlay (light leaks / dust & bokeh particles / film grain) — a semi-transparent layer with floating particles/light/easy fog over the video sequence, to make the frames look alive. Set 5-8 ready overlays to choose from; adjust transparency with a slider (0-100%); can also be applied to already finished video separately. Free overlays: Pexels, Pixabay, Mixkit, Videezy. Subtitles (embed or separate .srt). Clip processing for uniqueness. Resolution/format, video length, video share, search depth, number of candidates (default 5). LLM savings (set by default) Models — Claude (Anthropic API), separately for two tasks, parameter in settings: frame checking/selection (frequent operation) → Claude Haiku — cheap, analyzes still frames (images), for "suitable/not suitable" evaluation this is enough; script generation and breakdown (rare operation) → Claude Sonnet — for text quality. One API key for Claude gives access to all models (Haiku, Sonnet, etc.) — separate subscriptions are not needed. The model for each task can be changed in settings, without editing code. If no model is selected — the default model is used (does not break). Images are sent as previews of medium resolution (~512-768px on the long side) — sufficient for Claude to confidently recognize the content and assess its relevance to the scene. Do not send originals in full resolution (extra tokens), but also DO NOT reduce to the point of losing details — priority is on accuracy of assessment, savings are secondary. Preview resolution is a parameter in settings. Number of candidates per frame — parameter (default 5). Disk space management After assembly, all intermediate files (downloaded clips, temporary pieces) are automatically deleted — only the finished MP4 remains. The used database (list of clip IDs) is NOT included in auto-cleaning — this is a small text file, stored permanently. Only heavy media files are cleaned, not usage history. Cache limit (parameter, e.g. 5 GB): when exceeded, the oldest files are deleted. Folders for finished videos and temporary files are set in settings. Display of occupied space + "clear cache" button. Technical requirements Python. Modular structure (sources and LLM — interchangeable modules, to easily replace/add). All API keys are entered through the interface (section "Settings"), stored in the config, not in the code. Simple window (GUI — Tkinter/PyQt to choose), launched by double-click. All work is done through the interface (buttons, fields, sliders), without editing code. README (installation, keys, launch), clear logs, comments in the code. Interface (menu and control) Program with a side menu. Sections: New video — mode "generate by topic" OR "insert your script" (switch), voice selection, length, options, "Create" button. During assembly — progress screen by steps (script → voiceover → frame selection → assembly). My videos — list of finished videos. Voices — list of voices from ElevenLabs. Sources — list of all connected archives: shows added ones, can enable/disable each, set priority, and add your new source if you have its API key. Here also — source keys where registration is needed (Pexels, Pixabay, Europeana). Overlays — selection of atmospheric overlay and transparency. Settings — section "API keys" (Claude, ElevenLabs, etc. in one place), selection of Claude models, folders, resolution/format, cache limit. Clear cache — free up space. All keys and settings are entered through the interface, not through files/code. Result delivery All source code in open form (without obfuscation) + assembled working version. I can run it myself from the sources according to the instructions. The code is clean and understandable, so any other programmer can continue. All rights to the code after payment — mine. What I provide API keys (Claude, ElevenLabs, and sources where registration is needed). I pay for the paid ones myself. Examples of video references and topics for tests. Acceptance (ready if) I run → window → I enter the topic, choose a voice → "Create" → I receive a ready MP4. Video sequence according to the meaning of the text, alternation of video/photos, intro teaser 60 sec, voiceover on top. Works with at least 6 free sources, search across all at once + candidate selection. Frame selection: top-5 candidates from the code → selection of the best through Claude (for video — by still frames), limit 3 steps. Quality threshold depends on the source (archives — do not lose film frames).

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

A local Python script needs to be developed to automatically fill a Google Sheet with data from the company's internal service. Main logic: 1. Connect to the Google Sheet. 2. Find rows where the ID is filled but two target values are missing. 3. Form a link based on the template: https://internal-service.example/item/{ID} 4. Retrieve the two values (via API, if it exists, otherwise via Playwright). 5. Write the values back to the Google Sheet. 6. Mark the row as processed. 7. Continue processing the next rows. Requirements: • Python • Google Sheets API • Priority to use the official API • If no API — Playwright • No OCR, screen recognition, or mouse coordinates • Confidential data must not be logged • Configuration via .env • Test mode (without writing to the sheet) • Do not process already filled rows • Batch write changes to Google Sheets • Proper error handling and retries It is necessary to provide: - source code; - requirements.txt; - example .env.example; - installation instructions; - running instructions; - brief architecture description. Before starting implementation, please: 1. Suggest an architecture. 2. List the necessary accesses. 3. Ask clarifying questions. 4. Indicate the cost, deadlines, and estimated number of hours.

As part of enhancing the cybersecurity level of our infrastructure, we need to abandon the practice of storing "eternal" and static API keys, passwords, and integration tokens in the configuration files (.env, appsettings.json, config.yaml) of our microservices. Business Goal: Create a single secure storage point for confidential data (secrets) with a mechanism for their automatic updating (rotation) in external systems on a schedule. Our other services will request current tokens "on the fly" via API, which will minimize damage in case of compromise of any system component.Security Model and Encryption (Crypto Core) No secret should be stored in plaintext in the database. Upon application startup, a Master Key is passed to the environment variables. If the key is missing or has an invalid length, the service should fail at the initialization stage with a clear error in the logs. Each secret is encrypted before being written to the database using this Master Key. Upon request, it is decrypted in memory and returned in the response body.Audit Logging (Audit Trail) Any action with secrets (creation, reading by the service, successful or unsuccessful rotation) must be recorded in a separate log file audit.log (or a separate table in the database). Strict Taboo: It is strictly prohibited to record the actual values of secrets in the audit log (neither in plaintext nor in encrypted form).