File size: 31,416 Bytes
d065d19 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 | {"file_name": "01.webp", "id": "01", "category": "prose", "transcription": "WHITEBOARD OCR → FEW SAMPLE TEST\n\nI have rediscovered whiteboarding!\nBUT!\nIt's mounted at a really awkward height\nAND\nI have barely written by hand in years\nSo\nmy writing is hard to decipher", "description": "# Diagram Description\n\nA wall-mounted black-framed whiteboard photographed at a slight angle. A small black pen/eraser tray is attached to the lower-right corner. The writing is entirely handwritten text in black marker — no diagram shapes, arrows, or figures.\n\nThe content is a short handwritten note laid out as a title followed by a free-form paragraph. The title \"WHITEBOARD OCR → FEW SAMPLE TEST\" sits across the top, underlined, with a right-pointing arrow between \"OCR\" and \"FEW SAMPLE TEST\".\n\nBelow the title is a multi-line note written in a casual, cursive-leaning hand. Interjections (\"BUT!\", \"AND\", \"So\") are written on their own short lines between the main sentences, acting as connective beats. The overall tone is a self-deprecating preamble about the author's handwriting being difficult to read — effectively serving as the first test sample for the OCR experiment."}
{"file_name": "02.webp", "id": "02", "category": "list", "transcription": "SOME QUESTIONS\n\nCan you train a vision model / multimodal model on\na few reference images w/ manual transcriptions / ground\ntruths?\nWhat about abbreviations like w/ for with?", "description": "# Diagram Description\n\nA wall-mounted black-framed whiteboard shot from a slight low angle, with a pen tray visible at the bottom-right corner. The surface is otherwise blank — no drawn shapes, sketches, or diagrammatic elements. All content is handwritten text in black marker.\n\nAt the top-left is the underlined heading \"SOME QUESTIONS\". Beneath it are two handwritten questions written as flowing prose rather than as a bulleted list. The first spans three lines and asks whether a vision/multimodal model can be trained on a handful of reference images paired with manual transcriptions / ground truths. The second, on the fourth line, asks how the system should handle shorthand abbreviations, using \"w/\" (for \"with\") as the example.\n\nThe image functions as a meta-note about the OCR project itself: the whiteboard is both the subject and the medium of the question."}
{"file_name": "03.webp", "id": "03", "category": "list", "transcription": "SOME QUESTIONS\n\n• Can you train a vision model / multimodal model on\n a few reference images w/ manual transcriptions / ground\n truths?\n• What about abbreviations like w/ for with?\n• Could I get AI to create a custom TTF font & if so\n how much training data?", "description": "# Diagram Description\n\nThe same wall-mounted black-framed whiteboard as the previous frames, with the pen tray in the lower-right corner. No drawn shapes or figures — the content is entirely handwritten text in black marker, arranged as a titled bulleted list.\n\nThe heading \"SOME QUESTIONS\" is underlined at the top-left. Three bullet points follow, each marked with a solid dot on the left margin:\n\n1. A multi-line question about training a vision or multimodal model on a few reference images paired with manual transcriptions / ground truths.\n2. A shorter question about handling abbreviations such as \"w/\" for \"with\".\n3. A two-line question asking whether AI could generate a custom TTF font from the author's handwriting, and how much training data that would require.\n\nCompared to frame 02, this is an expanded version of the same note — bullets have been added and a third question appended. It reads as an in-progress brainstorming list on the feasibility of few-shot handwriting OCR."}
{"file_name": "04.webp", "id": "04", "category": "list", "transcription": "SOME QUESTIONS\n\n• Can you train a vision model / multimodal model on\n a few reference images w/ manual transcriptions / ground\n truths?\n• What about abbreviations like w/ for with?\n• Could I get AI to create a custom TTF font & if so\n how much training data?", "description": "# Diagram Description\n\nA near-duplicate of frame 03 — same black-framed wall-mounted whiteboard, same pen tray at the lower-right, same handwritten bullet list under the underlined heading \"SOME QUESTIONS\". There are no drawn shapes, arrows, or sketches; the image is pure handwritten text in black marker.\n\nThe three bullets cover: (1) training a vision/multimodal model on a few reference images with manual transcriptions/ground truths, (2) handling abbreviations like \"w/\" for \"with\", and (3) whether AI could generate a custom TTF font from the author's handwriting and how much training data that would need.\n\nThe photograph is taken from a slightly different angle and distance than frame 03 — likely a second capture of the same whiteboard state, useful as an additional sample of the same ground-truth text under different lighting/perspective conditions."}
{"file_name": "05.webp", "id": "05", "category": "list", "transcription": "Why / Applications\n\n○ Whiteboard workflows:\n ↳ Whiteboard → cleaned up tech diagram\n (nano banana). needs to be able to parse a\n \"OCR\" raw text. Objective: ↑ accuracy &\n ↓ pseudotext\n ↳ Whiteboard → text / todo list / SOL etc.", "description": "# Diagram Description\n\nThe wall-mounted black-framed whiteboard again, pen tray at the lower-right. The content is a mix of handwritten text and simple flow-style arrows — still no drawn shapes, but the arrows give it a light diagrammatic structure.\n\nAt the top-left is the underlined heading \"Why / Applications\". Below it is a single top-level bullet (\"○ Whiteboard workflows:\") with two indented sub-branches marked by hooked down-right arrows (\"↳\"), each describing a pipeline:\n\n1. **Whiteboard → cleaned-up tech diagram** (parenthetical note: \"nano banana\"). Annotation states the system needs to parse an \"OCR\" raw text pass, with objectives written as up/down arrows: ↑ accuracy and ↓ pseudotext (i.e. maximize accuracy, minimize hallucinated/garbled output).\n2. **Whiteboard → text / todo list / SQL etc.** — a second pipeline that converts whiteboard contents into structured textual artifacts such as todo lists or SQL.\n\nThe image documents two intended downstream use cases for the OCR model: diagram cleanup and structured-text extraction."}
{"file_name": "06.webp", "id": "06", "category": "diagram", "transcription": "Few Shot Learning - 1\n\n Apples\n ↗\n /\n( FRUITS ) → Oranges\n \\→ Bananas\n ↘\n Lemons\nKiwis ←\n ↙\nMelons", "description": "# Diagram Description\n\nA black-framed wall-mounted whiteboard with a pen tray in the lower-right corner. The drawing is a simple radial mind-map / spoke diagram rendered in black marker.\n\n- **Title** (top-left, underlined): \"Few Shot Learning - 1\".\n- **Central node**: the word \"FRUITS\" enclosed in a hand-drawn oval, positioned roughly in the middle-right of the board.\n- **Outgoing arrows**: from the central oval, five curved arrows radiate outward to individual fruit labels placed around it:\n - Upper-right: **Apples**\n - Right: **Oranges**\n - Lower-right: **Bananas**\n - Lower-center: **Lemons**\n - Left / lower-left: **Kiwis** and **Melons** (two arrows pointing leftward and down-left).\n\nThe arrows fan out in a classic \"hub and spokes\" pattern. No boxes or other structural elements — just the central labeled oval and the satellite labels. The diagram is clearly a teaching/demo example for a \"fruits\" concept, useful as a visual sample for the few-shot OCR training set."}
{"file_name": "07.webp", "id": "07", "category": "table", "transcription": "Few Shot Learning - 2\n\nWORLD CAPITALS\n\nCITY | COUNTRY\n--------------+---------\nDublin | Ireland\nAmman | Jordan\nWashington DC | USA", "description": "# Diagram Description\n\nBlack-framed whiteboard, close crop of the upper-left portion. The content is a simple hand-drawn two-column table in black marker.\n\n- **Title** (top-left, underlined): \"Few Shot Learning - 2\".\n- **Sub-heading** below the title: \"WORLD CAPITALS\" (all caps).\n- **Table structure**: drawn with a single vertical line (column separator) and a single horizontal line (header underline). No outer border.\n - **Column headers** (underlined): \"CITY\" on the left, \"COUNTRY\" on the right.\n - **Rows** (three, no row separators):\n 1. Dublin — Ireland\n 2. Amman — Jordan\n 3. Washington DC — USA\n\nThe diagram demonstrates a tabular layout for the OCR test set — a deliberate shift from prose/bulleted formats to a structured grid, so the model can be evaluated on parsing hand-drawn table geometry in addition to raw text."}
{"file_name": "08.webp", "id": "08", "category": "flowchart", "transcription": "Few Shot Learning - 3\n\n Zoom Meeting Prep SOP\n\n Is my hair a mess?\n / \\\n Y N\n ↓ ↓\n ┌─────────┐ ┌──────────┐\n │COMB HAIR│ │CHILL OUT!│\n └─────────┘ └──────────┘", "description": "# Diagram Description\n\nBlack-framed whiteboard, upper-left crop. The drawing is a simple hand-drawn decision flowchart in black marker, illustrating a binary yes/no branch.\n\n- **Title** (top-left, underlined): \"Few Shot Learning - 3\".\n- **Sub-heading**: \"Zoom Meeting Prep SOP\".\n- **Decision node**: the question \"Is my hair a mess?\" written as free text (no surrounding shape).\n- **Branching**: two arrows fan down from the question — one to the left labelled **Y** (yes), one to the right labelled **N** (no).\n- **Outcome nodes**: each branch terminates in a hand-drawn rectangular box containing an action:\n - **Y → [ COMB HAIR ]**\n - **N → [ CHILL OUT! ]**\n\nThis is the simplest possible flowchart: one decision, two outcomes. It adds boxed nodes and branching arrows to the OCR sample set, complementing the prose, bullet-list, mind-map, and table layouts used in earlier frames."}
{"file_name": "09.webp", "id": "09", "category": "list", "transcription": "Few Shot Learning - 3\n\nWhat about variability in how we write letters?\n ↳ Sometimes I dot i's. Sometimes I don't.\n ↳ Informal / missing punctuation?\n ↳ \"Code switching\" & misspellings?\n ↳ Sometimes I write 'b' like balloon & sometimes\n balloon (why? IDK!)", "description": "# Diagram Description\n\nBlack-framed wall-mounted whiteboard, pen tray in the lower-right corner. All content is handwritten text in black marker — no drawn shapes or figures, only small down-right hook arrows (\"↳\") used as sub-bullet markers.\n\n- **Title** (top-left, underlined): \"Few Shot Learning - 3\".\n- **Opening question** on the first line: \"What about variability in how we write letters?\"\n- **Indented sub-points**, each introduced by a \"↳\" arrow:\n 1. \"Sometimes I dot i's. Sometimes I don't.\"\n 2. \"Informal / missing punctuation?\"\n 3. \"'Code switching' & misspellings?\"\n 4. A two-line observation that the author sometimes writes the letter **'b'** one way and sometimes differently (\"balloon\" vs \"balloon\"), ending with a parenthetical \"(why? IDK!)\".\n\nThe note catalogues sources of intra-writer variability that the few-shot OCR system will need to handle — letter-form drift, inconsistent punctuation, informal register, and shorthand/misspellings. It is a meta-note about the difficulty of the OCR task rather than a structured diagram."}
{"file_name": "10.webp", "id": "10", "category": "list", "transcription": "Few Shot Learning - 3\n\nWhat about txt like speech?\n ↓\n\nLLMs R cool → LLMs are cool!\n or incorrect uppercase\nLLMs R cool → LLMs are cool", "description": "# Diagram Description\n\nBlack-framed whiteboard, upper-left crop taken at an angle. All content is handwritten text in black marker, laid out as two input-to-output transformation examples connected by arrows.\n\n- **Title** (top-left, underlined): \"Few Shot Learning - 3\".\n- **Prompt line**: \"What about txt like speech?\" with a single down arrow underneath, acting as a visual \"so, consider:\" pointer.\n- **Transformation examples**, each a before → after pair:\n 1. **\"LLMs R cool\"** → **\"LLMs are cool!\"** (expanding the text-speak \"R\" to \"are\" and adding an exclamation mark).\n 2. Between the two examples is a connecting note: **\"or incorrect uppercase\"**.\n 3. **\"LLMs R cool\"** (with the \"M\" underlined to flag an uppercase issue) → **\"LLMs are cool\"** (with the \"M\" underlined in the output as well).\n\nThe image illustrates how the OCR / post-processing layer should normalise \"txt-speak\" and correct casing — two failure modes the author wants the few-shot model to handle. It uses arrows to make the input→output mapping explicit, giving it a light diagrammatic structure."}
{"file_name": "11.webp", "id": "11", "category": "diagram", "transcription": "PERSONALISED AI → BASIC IDEA\n\nUse-case: Human wants to create a task-specific AI agent\n& frontload context\n\nTactic: Create AI agent to \"interview\" you → transcribe\n (could be voice → voice!) → extract content → vector DB\n\n[bot] ──> [ What's your fav food? ]\n [ PIZZA! ] :)\n ↓\nCREATE MEMORY\n┌───────────────┐\n│ Daniel's fav │\n│ food = PIZZA │\n└───────────────┘\n ↓\n ( VECTOR DB )", "description": "# Diagram Description\n\nWide shot of the black-framed whiteboard with a hand-drawn concept diagram in black marker. The layout combines headed text, prose annotations, and a small pictographic flow at the bottom.\n\n- **Title** (top, underlined): \"PERSONALISED AI → BASIC IDEA\".\n- **Use-case line**: describes a human wanting to create a task-specific AI agent and \"frontload context\".\n- **Tactic line** (prose with inline arrows): \"Create AI agent to 'interview' you → transcribe → extract content → vector DB\", with a parenthetical note that the interview \"could be voice → voice!\".\n- **Pictographic flow** (bottom half):\n - A small **bot face icon** (square head with smiley) connected by arrows to two **speech-bubble rectangles**: the first says \"What's your fav food?\", the second (reply) says \"PIZZA!\", with a small smiley emoticon beside it.\n - An arrow labelled **CREATE MEMORY** drops down from the bot to a rectangular memory record: **\"Daniel's fav food = PIZZA\"**.\n - A final downward arrow leads from the memory record into a drawn cylinder (database symbol) labelled **VECTOR DB**.\n\nThe diagram depicts an end-to-end pipeline: interview → transcribe → extract structured memory → persist to a vector store, illustrated with a conversational exchange."}
{"file_name": "12.webp", "id": "12", "category": "diagram", "transcription": "(could be voice → voice!) → Extract content → V[ector DB]\n\n[bot] ──> [ What's your fav food? ]\n [ PIZZA! ] :)\n ↓\nCREATE MEMORY\n┌───────────────┐\n│ Daniel's fav │\n│ food = PIZZA │\n└───────────────┘\n ↓\n ( VECTOR DB )", "description": "# Diagram Description\n\nA close-up of the lower portion of the same \"Personalised AI — Basic Idea\" whiteboard shown in frame 11. The top edge of the crop clips the tactic line (\"(could be voice → voice!) → Extract content → V[ector DB]\"), so only the tail of that text is visible. The focus is the pictographic flow.\n\n- **Bot icon**: a small square face with a smiley, representing the AI agent.\n- **Speech bubbles** (two stacked rectangles to the right of the bot):\n - Question: \"What's your fav food?\"\n - Answer: \"PIZZA!\" with a circular smiley emoticon beside it.\n- **\"CREATE MEMORY\"** label with a downward arrow from the bot.\n- **Memory record** — a rectangular box containing \"Daniel's fav food = PIZZA\".\n- **Vector DB** — a downward arrow from the memory box leads into a hand-drawn cylinder labelled \"VECTOR DB\".\n\nThis frame is essentially a zoom-in of the bottom half of frame 11, useful as a second OCR sample of the same drawn pipeline at higher detail."}
{"file_name": "13.webp", "id": "13", "category": "diagram", "transcription": "PERSONALISED AI — WORKFLOW 2\n\n(user) Hey bot! Let's build out my food & drink\n context store!\n\n(bot) Sure! Let me see what gaps we have...\n ↓\n READ → (vector DB) → (bot) Hmm. nothing about\n beer preferences!\n\n(bot) What beer do you prefer?\n(user) Super bitter IPAs CREATE →\n ┌─────────────────────┐\n (bot) │ User's fav beer = │\n │ bitter IPA │\n └─────────────────────┘", "description": "# Diagram Description\n\nBlack-framed whiteboard showing a hand-drawn conversational workflow diagram in black marker. Pen tray visible at the lower-right corner.\n\n- **Title** (top, underlined): \"PERSONALISED AI — WORKFLOW 2\".\n- **Characters**: two small drawn head icons are used throughout as speaker avatars — a round smiling face represents the **user**, a square head with a neutral/slightly-worried face represents the **bot**.\n- **Dialogue flow** (top half):\n - User avatar: \"Hey bot! Let's build out my food & drink context store!\"\n - Bot avatar: \"Sure! Let me see what gaps we have...\"\n - A downward arrow labelled **READ** leads from the bot to a small drawn cylinder (**vector DB**), then an arrow curves back up to the bot, who replies: \"Hmm. nothing about beer preferences!\"\n- **Second exchange** (bottom half):\n - Bot: \"What beer do you prefer?\"\n - User: \"Super bitter IPAs\"\n - An arrow labelled **CREATE** points to a boxed memory record containing **\"User's fav beer = bitter IPA\"** (with the bot avatar on the left of the box).\n\nThe diagram depicts the \"context-gap-detection\" loop of a personalised AI: read the vector DB → identify missing fields → ask the user → write the new memory back."}
{"file_name": "14.webp", "id": "14", "category": "diagram", "transcription": "PERSONALISED AI — LATENT VALUE AREAS\n\n (user) (bot)\n USER LLM'S\n PROMPTS OUTPUTS\n │ │\n ↓ ↓\n PROMPT → CONVO (1)\n HISTORY (2) STORE\n ↓ │ ↘\n ↓ ↓ WIKI\n PROMPT DOWNSTREAM\n LIB → USER OUTPUTS\n CONTEXT ↙ ↓ ↘\n PODCASTS CPD DOCS", "description": "# Diagram Description\n\nBlack-framed whiteboard with a hand-drawn branching tree / flow diagram in black marker. Pen tray at the lower-right. The diagram is split into two major trunks originating from two drawn head icons at the top.\n\n- **Title** (top, underlined): \"PERSONALISED AI — LATENT VALUE AREAS\".\n- **Two source nodes at the top** (drawn as cartoon head icons):\n - Left: a round smiley face labelled **USER** with downstream label **PROMPTS**.\n - Right: a square head labelled **LLM'S** with downstream label **OUTPUTS**.\n- **Left branch (from USER → PROMPTS)**:\n - Arrow down to **PROMPT HISTORY** (tagged with a circled \"2\").\n - From there, an arrow down to **PROMPT LIB** and a second arrow to **USER CONTEXT**.\n- **Right branch (from LLM'S OUTPUTS)**:\n - Arrow to **CONVO STORE** (tagged with a circled \"1\").\n - From CONVO STORE two arrows fan out: one to **WIKI** (upper-right), one down to **DOWNSTREAM OUTPUTS**.\n - **DOWNSTREAM OUTPUTS** fans into three leaves: **PODCASTS**, **CPD**, and **DOCS**.\n\nThe diagram maps out under-exploited reservoirs of personal context generated during AI use — prompt history on the user side, conversation logs and their derivatives on the model-output side — and proposes downstream artifacts (libraries, wikis, podcasts, CPD records, docs) that can be built from each stream."}
{"file_name": "15.webp", "id": "15", "category": "diagram", "transcription": "(bot) LLM'S OUTPUTS → CONVO (1)\n STORE\n │ ↘\n ↓ WIKI\n DOWNSTREAM\n OUTPUTS\n ↙ ↓ ↘\n PODCASTS CPD DOCS", "description": "# Diagram Description\n\nA close-up crop of the right-hand branch of the \"Personalised AI — Latent Value Areas\" diagram from frame 14. The title and left-side branch are out of frame; only the LLM-outputs sub-tree is visible.\n\n- **Top-left**: a drawn bot head icon labelled **LLM'S OUTPUTS**, with a left-edge fragment of the user branch (\"...XT\" — presumably \"USER CONTEXT\") clipped at the margin.\n- An arrow from LLM'S OUTPUTS leads right to **CONVO STORE**, tagged with a circled \"1\".\n- From CONVO STORE two arrows fan out:\n - Up-right to **WIKI**.\n - Down to **DOWNSTREAM OUTPUTS**.\n- **DOWNSTREAM OUTPUTS** fans into three leaves with arrows:\n - **PODCASTS** (lower-left)\n - **CPD** (centre)\n - **DOCS** (right)\n\nThis is a higher-detail view of the right half of frame 14 — useful as an additional OCR sample of the same sub-tree at larger scale."}
{"file_name": "16.webp", "id": "16", "category": "diagram", "transcription": "(user) (bot)\nUSER LLM'S\nPROMPTS OUT[PUTS]\n │\n ↓\nPROMPT ────┐\nHISTORY (2) │\n │ ↓\n ↓ USER\nPROMPT LIB CONTEXT", "description": "# Diagram Description\n\nA close-up crop of the left-hand branch of the \"Personalised AI — Latent Value Areas\" diagram from frame 14. The title is out of frame; only the user-prompts sub-tree is visible, with the right-side LLM branch partially clipped at the right edge.\n\n- **Top**: a drawn smiley square head icon labelled **USER**, with label **PROMPTS** below.\n- **Right-edge fragment**: the square bot head icon labelled **LLM'S** with partial **OUT...** (PUTS) label, clipped.\n- Arrow from USER/PROMPTS curves down to **PROMPT HISTORY**, tagged with a circled \"2\".\n- From PROMPT HISTORY, two arrows:\n - Down to **PROMPT LIB**.\n - Right-then-down to **USER CONTEXT**.\n\nThis is a higher-detail view of the left half of frame 14 — useful as an additional OCR sample of the same sub-tree at larger scale."}
{"file_name": "17.webp", "id": "17", "category": "prose", "transcription": "Other Common Whiteboard Imperfections\n\n o Use X or strikethrough\n\n Here's cool ~~ideas~~\n ✗ ideas\n are\n some", "description": "# Diagram Description\n\nBlack-framed whiteboard with handwritten notes in black marker illustrating common editing/imperfection conventions on a whiteboard.\n\n- **Heading** (top, not underlined): \"Other Common Whiteboard Imperfections\".\n- **Bullet** (single \"o\" marker): \"Use X or strikethrough\".\n- **Example sentence** below shows a live mid-edit: \"Here's [X over 'a'] cool ~~ideas~~ ideas\" where the word \"a\" has an X struck through it, \"ideas\" (the first instance) is struck through with a line, and a replacement \"ideas\" is written to the right.\n- Below the main line, two caret-inserted words — **are** and **some** — are squeezed in to signal an intended rewrite: \"Here's some cool ideas\" → \"Here are some cool ideas\".\n\nThe board demonstrates the kind of scribble-style corrections (X-out letters, strikethrough words, caret insertions) that an OCR system needs to handle when interpreting hand-edited whiteboard content."}
{"file_name": "18.webp", "id": "18", "category": "table", "transcription": "Task: Whiteboard → Calendar\n\nMy Week Plan\n\n SUN MON TUES WEDS THURS FRI SAT\n ─────────────────────────────────────────────────\n ACCOUNTANT\n\n0900\n\n1200 DAY\n OFF\n1500 DOCTOR\n\n1800\n\n2100\n\n0000", "description": "# Diagram Description\n\nBlack-framed whiteboard depicting a hand-drawn weekly calendar/timetable grid, framed as an OCR-to-calendar conversion task. Pen tray visible at the lower-right corner.\n\n- **Title** (top): \"Task: Whiteboard → Calendar\".\n- **Sub-title**: \"My Week Plan\".\n- **Column headers** (underlined horizontal line beneath): **SUN, MON, TUES, WEDS, THURS, FRI, SAT** — days of the week across the top.\n- **Row labels** (left column, time slots, 24-hour format): **0900, 1200, 1500, 1800, 2100, 0000**.\n- **Entries** populated in the grid:\n - **TUES** column, top: **ACCOUNTANT** (likely a morning appointment).\n - **SUN** at 1500: **DOCTOR**.\n - **THURS** around 1200: **DAY OFF**.\n\nThe diagram frames a realistic capture scenario: a user writes a weekly plan on a whiteboard, and the OCR pipeline should parse the grid into structured calendar events (day, time, label)."}
{"file_name": "19.webp", "id": "19", "category": "mixed", "transcription": "Task: Whiteboard → Mixed Elements.\n\nWhiteboard → Handwriting │ Questions\n ↳ Gather samples │ ─────────\n ↳ Determine priorities │ Existing datasets?\n ↳ Send to David │ Separate fine tunes for\n │ handwriting creation vs.\n │ reading?", "description": "# Diagram Description\n\nBlack-framed whiteboard split into two columns by a vertical divider line, sketched in black marker. Pen tray visible at the lower-right. The board captures a task breakdown plus an open-questions panel.\n\n- **Title** (top): \"Task: Whiteboard → Mixed Elements.\" — framing this as an OCR task about boards that contain a mix of handwriting, diagrams, and structure.\n- **Left column** — headed \"Whiteboard → Handwriting\", with three nested sub-bullets drawn as ↳ arrows:\n - **Gather samples**\n - **Determine priorities**\n - **Send to David**\n- **Right column** — headed (underlined) \"Questions\":\n - \"Existing datasets?\"\n - \"Separate fine tunes for handwriting creation vs. reading?\"\n\nThe diagram presents a working plan for building the handwriting/OCR dataset (this very repo) alongside unresolved research questions about dataset availability and whether generation and recognition should be modelled as separate fine-tunes."}
{"file_name": "20.webp", "id": "20", "category": "diagram", "transcription": "Geopol Predictions: Current Model A_x = Agent X\n Sequential\n SA_x = Subagent\nA_1 → Job\n Actuation &\n Mgmt RSS EXA AI\n ↘ ↙\n Google\n News\n ↘\n A_2 SITREP Gen\n Grounding ┌──────┐\n │ LLM │\n ↘ → SA_1 ←│COUNCIL│\n A_3 Prediction └──────┘\n MGR → SA_2\n Pass SITREP, define\n task, structured output\n\n A_4 Analysis, Charts, Comparison w/\n Run X-1\n\n A_5 Generate Report (Typst)\n ↓\n CONCAT → PDF", "description": "# Diagram Description\n\nBlack-framed whiteboard with a dense hand-drawn multi-agent pipeline diagram in black marker. Pen tray visible at the lower-right corner.\n\n- **Title** (top-left): \"Geopol Predictions: Current Model\".\n- **Legend** (top-right): \"A_x = Agent X\", \"Sequential\", \"SA_x = Subagent\".\n- **Agent chain** (left side, top-to-bottom):\n - **A_1 — Job Actuation & Mgmt**: the entry-point agent.\n - **A_2 — SITREP Gen / Grounding**: generates a situation report, grounded in external feeds. Upstream inputs drawn above: **RSS** and **EXA AI** both feeding into **Google News**, which flows into A_2.\n - **A_3 — Prediction MGR**: \"Pass SITREP, define task, structured output\". Fans out to two subagents: **SA_1** (routed through/via an **LLM COUNCIL** boxed node on the right) and **SA_2**.\n - **A_4 — Analysis, Charts, Comparison w/ Run X-1**: diff-vs-prior-run analysis stage.\n - **A_5 — Generate Report (Typst)**: renders the final report using Typst.\n- **Output path** (bottom): a curving arrow from the agent chain down to **CONCAT** → **PDF** — concatenation into the final PDF deliverable.\n\nThe diagram maps the current architecture of a geopolitical-predictions agent pipeline: ingest news feeds → generate SITREP → dispatch prediction tasks to an LLM council of subagents → analyse against previous run → typeset and concatenate into a PDF report."}
{"file_name": "21.webp", "id": "21", "category": "diagram", "transcription": "Whiteboard → Dev Spec → Agent Dev\n\nWhiteboard(s) ─────────→ Gemini API\n\nGemini ─→ Cleaned Tech Diagrams (image 2 image)\n ↘ Narrative of Agent Model (image 2 txt)\n\nAgent 2: Convert these things → Dev Spec\n ↳ Plan def w/ human\n ↳ Generate tasks, define subagents,\n checkpoints etc", "description": "# Diagram Description\n\nBlack-framed whiteboard with a hand-drawn pipeline sketch in black marker. Pen tray visible at the lower-right corner. The board maps a two-stage agent workflow that turns whiteboard photos into a structured development specification.\n\n- **Title** (top): \"Whiteboard → Dev Spec → Agent Dev\".\n- **Stage 1 — ingestion**: \"Whiteboard(s) ───→ Gemini API\" — raw whiteboard images fed to the Gemini API.\n- **Stage 1 outputs** (from Gemini, two branches):\n - **Cleaned Tech Diagrams** (labelled *image 2 image*) — the vision model produces a cleaned/rendered diagram from the photo.\n - **Narrative of Agent Model** (labelled *image 2 txt*) — the vision model produces a prose description of the depicted agent architecture.\n- **Stage 2 — synthesis (Agent 2)**: \"Convert these things → Dev Spec\", with two nested sub-bullets drawn as ↳ arrows:\n - **Plan def w/ human** — the agent co-defines the plan with a human in the loop.\n - **Generate tasks, define subagents, checkpoints etc** — produces the concrete task list, subagent roster, and milestone checkpoints.\n\nThe diagram describes a whiteboard-to-code meta-pipeline: photograph a system design on a whiteboard, let Gemini produce both a cleaned diagram and a narrative, then have a second agent synthesise those artifacts into an executable development specification."}
|