455 commits · +242,969 lines
A portfolio archive put 45% of its documents in the unknown bucket and classified a cap table as product documentation. A shared text layer, a name-stripping classifier, and the legal extractor that was never built take that to 1%.
Michael StatonThe deck analyst batches five slides per call and merges them, which destroys slide identity by design. A second agent keeps the slide as the unit: render, transcribe verbatim, emit typed content, one markdown file each — plus date resolution, deck lineage, and a renamer that puts a portfolio archive on convention.
Michael StatonTwo collectors run over transcribed slides and isolate the parts a reader has to see: charts with their axes and legends intact, and any visual that communicates a concept. Decoration is named and skipped. A deck-level index makes the result reusable without opening a single slide document.
Michael StatonSonar returns the sources it really fetched on every call. We read none of them. 37% of a memo's citations were invented, and a third of the rest pointed at real pages with the wrong title.
Michael StatonPreviews fetched a page, showed it, and threw it away. Now they file it. Found a bug on the way that was quietly stripping `internal_only` from real client sources.
Michael StatonEvery gate in the pipeline asked whether a URL resolves; none asked whether it was approved. That single missing predicate is why the researcher, enricher, and validator have been fighting each other for months — a fabricated URL that happens to be live passes every reachability check ever written. This lands the membership predicate in both cleanup gates, no-ops citation enrichment in codified mode, strips out-of-set sources from fact corrections, adds an LLM-free SearXNG candidate seeder, and exposes six sidecar endpoints for the approval surface. Also fixes a pre-existing bug that was carrying every dropped URL into the final draft.
Michael StatonThe last six days surfaced enough orchestrator-level work to log: a new agent-skill that codifies the end-to-end procedure for onboarding a VC firm (setup-new-firm), two CLI scripts for capturing DocSend decks programmatically, three issue-resolution write-ups naming real pipeline gaps the recent runs exposed (competitive research not flowing into prose, missing exit-comp agent, unsolicited investor verdicts), one new spec for per-deal analyst focal points, and a small skills-directory rename (skills/ → agent-skills/) to match the parent repo's convention. Also adds a changelog-scope discipline to CLAUDE.md so firm-private content stays out of the orchestrator's log.
Michael StatonAlpha JWC Ventures becomes the sixth firm wired into the orchestrator under the established io/<firm>/ private-submodule pattern. This is a structural addition only — registering the submodule path in .gitmodules and pinning the initial commit. The firm-private content (brand config, customized 9-section 7Cs outline, first-deal setup) lives in the alpha-jwc-secure repo and is documented in its own changelog there, per the CLAUDE.md scope rule.
Michael StatonCitation assembly was rewriting section files. Section files now stay the durable source of truth and keep their semantic IDs.
Michael StatonThree reproducible citation-assembly bugs squashed.
Michael StatonYesterday's v0.0.10 run made a problem legible: 33 hand-curated URLs went into the writer, 7 came out cited in the prose, 26 sat untouched. The fix wanted to be a new agent, and the first instinct was to name it 'Citation Diversity Enricher' and slot it between the writer and the existing 'cite' node. Today's planning work landed somewhere different — the existing cite node runs upstream of the writer (not after), the new agent's job is genuinely orthogonal to enrichment, and the right name is Citation Coverage Promoter. No code shipped; a plan got corrected before it generated a confused agent.
Michael StatonFor months MemoPop's citation validator quietly skipped its URL accessibility check on almost every citation — because a date-format warning was enough to disable it. Two small changes put it back online, and upgrade the underlying check from a HEAD-only 'did this URL respond' to a GET-with-body-sniff that catches soft-404s and paywall stubs the prior pass had been rubber-stamping as valid sources.
Michael StatonThree things shipped today that together put the human analyst back in the middle of the citation pipeline: URL-keyed validation so the validator finally sees every source (not just the 6 that survived a per-file numbering collision), a Sources.md codified-mode loader that lets the writer compose only from analyst-approved URLs, and a new aggregate_sources agent that halts the pipeline mid-run with a curated draft so the analyst can rank and prune before any prose gets written.
Michael StatonA new Firecrawl-driven pipeline walks any organization's website, finds the team page, extracts a structured roster, and back-fills missing socials by name. Phase 2 of Team-and-People-Metadata-Ingestion ships.
Michael StatonEnd-to-end integration of the desktop client with the orchestrator's FastAPI sidecar across four phases: pytest coverage on the server, a Rust transport layer with lazy spawn and /healthz poll, a TypeScript Transport implementation with SSE event subscription, and a Generate flow that streams milestones live into the UI.
The sidecar gains structured milestone events extracted from agent prints, and a missing scaffold step that was sending firm-scoped runs to legacy output/.
Michael StatonFirst contact with the desktop client surfaced four ways the orchestrator-as-a-service could go sideways. All four are closed in this pass.
Michael StatonA new brand_fetch module drives a Claude tool-use loop over a firm's homepage, and two endpoints expose it. The HTTP surface keeps growing; agents stay untouched.
Michael StatonThe investment memo orchestrator has been a 30-agent Python CLI for months — you type a company name, wait 15 to 45 minutes, and a fully cited 50-file deal memo lands in `output/`. It just works. Now it also speaks HTTP.
New guided CLI application that walks users through memo generation, export, and iteration without requiring knowledge of individual commands, flags, or file paths.
The Table of Contents was consistently missing from final memo output despite the `toc_generator` agent running successfully mid-pipeline. Two separate ordering bugs caused the TOC to either never generate or get destroyed before finalization.
Introduces a three-step LLM-powered fact-check pipeline (extract → verify → correct) and a per-section comprehensive source catalog. The mechanical-only fact-checker is upgraded to use Perplexity Sonar Pro for independent claim verification and Claude for surgical corrections.
Strengthens entity disambiguation to prevent research agents from conflating similarly-named organizations. Adds both improved prompt language and a mechanical post-synthesis override that forces correct company identity regardless of what the LLM returns.
New pipeline agent that mechanically enforces correct markdown citation spacing — no LLM, pure regex. Fixes the recurring problem of clumped, duplicated, and improperly spaced citations that affect Obsidian rendering, HTML/PDF exports, and readability.
New pipeline agent and standalone CLI tool that distills the full investment memo into a single-page visual summary. Serves as both a standalone deal teaser and a cover sheet for the full memo export.
Consolidated all final draft file resolution into `src/final_draft.py` and bumped the canonical prefix from `6` to `7`. Previously, three agents each had their own inline glob logic to locate the final draft — now they all call shared utility functions.
New standalone CLI for running the table generator agent outside the full pipeline. Scans memo sections and structured state data for tabular data opportunities, generates markdown tables, and inserts them into section files.
Rewrote the deck screenshot classification, naming, and injection system to fix two core problems: (1) Claude Vision was misclassifying slides because it only had 7 vague categories to choose from, and (2) the injection agent was spamming sections with duplicate, poorly-placed im…
Two new agents that discover, evaluate, and classify competitors using Perplexity Sonar Pro. The system splits competitive analysis into two phases with opposing cognitive biases: the researcher uses **completion bias** to find as many candidates as possible, while the evaluator…
A complete standalone module for parsing market research PDFs (McKinsey, CB Insights, PitchBook, Gartner, etc.) into markdown with properly formatted Obsidian-style citations.
Fixes to the citation pipeline addressing two critical issues: (1) the writer agent's regex only matched numeric citation keys (`[^1]`) and missed alphanumeric keys (`[^deck]`), causing citation loss during polishing; and (2) the citation enrichment agent was appending duplicate…
Critical fixes to the citation pipeline that were causing citation definitions (the source references at the bottom of files) to be lost during processing.
This release fixes an issue where the 12Ps Scorecard Summary section (section 08) would contain LLM "excuses" instead of actual scorecard data.
This release refactors final draft operations into a single centralized module (`src/final_draft.py`) and updates the artifact numbering scheme.
When resuming a workflow with `cli/resume_from_interruption.py` for firm-scoped deals, the writer agent would use legacy `output/` paths instead of firm-scoped `io/{firm}/deals/{deal}/outputs/` paths.
Fixed a critical issue where Perplexity was returning garbage/meta-commentary for non-standard section names like "Opening" and "Organization" in the 12Ps outline.
This release adds automatic 12Ps scorecard integration to the memo workflow. Previously, the scorecard evaluator generated detailed scores in `5-scorecard/12Ps-scorecard.
This release adds explicit wrong-entity exclusion to the disambiguation system. Deal configurations can now specify a `disambiguation` array of URLs/domains that belong to **different companies** with similar names.
This release completes the firm-scoped IO system, enabling private, firm-specific configurations, inputs, outputs, and exports. The system allows each firm (e.g.
Added a comprehensive scorecard evaluation system that scores completed investment memos against a structured 12-dimension framework (12Ps: Persona, Pain, Proposition, Problem, Possibility, Positioning, People, Process, Product, Potential, Progress, Plan).
Began major refactor to reorganize the flat `data/` and `output/` directories into a hierarchical `io/` directory organized by firm and deal. This enables private git submodules for firm-specific data while keeping the orchestrator open-source.
Completed the firm-scoped IO path resolution system, enabling deals to be stored in `io/{firm}/deals/{deal}/` with automatic fallback to the legacy `data/` and `output/` directories. All agents now pass the `firm` parameter through the workflow.
Added a focused set of CLI tools to tighten the memo workflow end-to-end:
A memo for Sava, the YC trust-administration startup, confidently reported a $19M Series A led by Balderton — except that round belonged to Sava Technologies, a biosensor spinout. The data was real and well-cited, just about a different company.
Three CLI tools each had their own way of stitching a memo back together, so citation numbers collided and the table of contents went stale. Now there is one assembler, and everything else delegates to it.
Three new CLI tools let you repair a finished memo without regenerating it: add citations to prose you already like, hyperlink every organization it mentions, or score the whole thing and see the breakdown.
Added explicit "Venture Capital mindset" framing to all memo-writing agent prompts to ensure generated content focuses on opportunity and upside rather than defaulting to Private Equity-style skepticism.
Created a comprehensive outline for evaluating LP commitments to emerging manager VC funds, integrating Hypernova Capital's proprietary 12-dimension GP evaluation framework directly into the memo generation system.
Implemented the Scorecard Agent and CLI tool to automatically generate Hypernova Capital's 12-dimension scorecards for emerging manager fund commitment memos. This completes the scorecard system introduced in changelog 2025-11-27_01.
Introduced a new multi-agent dataroom analyzer system for processing investment datarooms, plus improvements to trademark handling and export versioning.
Implemented Phase 2 of the Dataroom Analyzer: the Competitive Extractor. This extracts and synthesizes competitive landscape data from battlecard PDFs and outputs structured analysis as separate, numbered artifact files for better manageability.
Added two new extractors to the Dataroom Analyzer: Cap Table Extractor and Financial Extractor. These complete the core Phase 2 extractors for investment dataroom analysis.
Added the Traction Extractor to complete Phase 2 extractors (4 of 5). This extractor pulls customer metrics, revenue data, pipeline information, and partnerships from pitch decks and other traction-related documents.
Added the Team Extractor, completing all Phase 2 extractors (5/5). The dataroom analyzer now extracts comprehensive structured data from investment datarooms.
Added a new TOC Generator agent that automatically creates a Table of Contents with working anchor links for the final investment memo. The TOC includes main sections (h2) and subsections (h3) that function in both HTML and PDF exports.
Added the Synthesizer module to cross-reference data across all extractors, detect conflicts, identify data gaps, and generate a unified view of company metrics with confidence scores.
Major refactoring to organize 38 scattered Python files and shell scripts into logical directories, improving developer experience and project maintainability.
Implemented checkpoint-based resume functionality to recover from workflow interruptions without losing progress or re-running expensive API calls. System now detects completed agents from saved artifacts and resumes execution from the last successful checkpoint.
Enhanced PDF export quality with proper logo rendering, Unicode symbol support, and GHCP Partners brand configuration. Fixed critical rendering issues that caused logo distortion and citation symbols to display incorrectly in WeasyPrint-generated PDFs.
Implemented comprehensive prompt engineering fixes to eliminate LLM fabrication of pricing, revenue, traction, and financial metrics in investment memos.
Integrated Perplexity's `@source` syntax into the research workflow to ensure high-quality data from premium sources like @crunchbase, @pitchbook, @statista, and @cbinsights.
Implemented the fact-checker agent to prevent hallucinated metrics and claims in investment memos. This agent verifies all factual claims against research sources, catching fabricated revenue figures, pricing data, customer counts, and other metrics before final memo assembly.
Complete implementation of image-based PDF deck analysis using Claude's Vision API with intelligent batch processing. This solves the recurring issue of Docsend-exported decks that contain only images with no extractable text.
Complete integration of Perplexity Sonar Pro for section-specific research WITH citations as a core part of the memo generation pipeline. This fundamentally changes the architecture from "write then add citations" to "research with citations then polish.
Complete refactor of the enrichment pipeline to process memo sections individually rather than as a single monolithic document. This eliminates API timeout issues and enables proper citation management with Perplexity Sonar Pro.
Major architectural improvements to prevent API timeouts through section-by-section processing, plus comprehensive brand expansion with MaC VC assets and SVG logo support.
This session focused on building and refining a comprehensive export/conversion system for investment memos, enabling professional PDF generation with Hypernova Capital branding.
Implemented a comprehensive multi-brand export system that allows managing and exporting memos for multiple VC firm clients from a single installation.
Solved a critical PDF rendering issue where page backgrounds didn't extend to all edges and content lacked proper per-page spacing. Discovered WeasyPrint's `@page background-color` property enables edge-to-edge colored backgrounds while maintaining proper content margins on every…
Implemented automatic citation consolidation in HTML exports to match Obsidian's footnote behavior. When Perplexity AI cites the same source multiple times using `[^2]` throughout the markdown, Pandoc was incorrectly converting each instance into a separate numbered footnote, res…
Four enrichment agents now run after the draft is finished — links, visualizations, socials, citations — turning plain prose into a memo with 19 organization hyperlinks and footnotes that actually resolve in Obsidian.
Historical entry — Core Multi-Agent System, Templates & Style Guides, CLI Interface.
Five specialized agents under a LangGraph supervisor, wired to live web search — the Week 1 proof of concept wrote a real investment memo on Aalo Atomics and scored 7.5 out of 10 against its own validator.
Citations finally work. A new enrichment agent slots in between the writer and the validator, and every run now leaves a 16-file artifact trail you can audit step by step.
The pitch deck now goes in first: a new agent reads the PDF before any web research runs, because proprietary beats whatever the internet knows. Separate templates for direct versus fund investments landed alongside it.