Treat extraction quality as data
Native parsing remained the fast path, but glyph and quality checks decided when a page needed selective or full OCR instead of silently accepting garbage.
AtlasConnect → Pitchfire
Private production work · 2025
From incoming pitch deck to an investment decision.
The system turned messy submissions into structured opportunities, firm-specific research, and a decision workflow people could actually operate.
This case study covers the workflow and engineering boundaries I can discuss publicly. It omits customer submissions, private prompts, scoring weights, research instructions, and private source code.
01 · System flow
One pipeline from raw submission to firm decision.
The interesting part was not any single model call. It was making messy intake, document recovery, research, structured evaluation, and ordinary collaboration behave like one product.
- 01Intake
- 02Extraction
- 03Enrichment and research
- 04Firm decision
01 · Intake
Accept the opportunity through the channel it actually arrived in.
Embedded forms, direct uploads, DocSend imports, and an AWS Lambda email-intake path I built and operated all converged on the same opportunity workflow.
Multiple entry points mattered because a venture workflow cannot require every founder, analyst, or forwarded deck to begin inside one idealized form.
02 · Extraction
Make bad documents explicit instead of trusting the happy path.
The ingestion layer handled PDF, DOCX, PPTX, images, and text. It attempted native extraction first, detected glyph and quality failures, then used selective or full OCR fallback with parallel page work when the source required it.
Structured opportunity fields and visible progress turned that document work into an inspectable product step rather than a blocking upload spinner.
03 · Fit and research
Build firm context and opportunity context before asking for a verdict.
Firm and opportunity preparation ran concurrently, then fed synthesis, thesis-fit evaluation, market research, and leadership research. Bounded navigation and background execution kept research useful without giving it an unbounded request lifecycle.
The public boundary here is the staged system; private scoring prompts, weights, and research instructions stay out of scope.
04 · Firm workflow
A decision product still needs the ordinary product work.
Opportunity stages, voting, requests, discussions, reminders, notifications, and realtime completion feedback carried the enriched deal through the firm's human decision process.
That workflow is what separated a production product from a collection of LLM calls that happened to read pitch decks.
03 · Engineering choices
The pipeline stayed observable at every expensive boundary.
Separate preparation from judgment
Firm and opportunity analyses could run concurrently before synthesis, keeping expensive research observable and preventing one opaque prompt from owning the whole decision.
Move slow work off the request path
Background execution and realtime progress let ingestion, research, and memo work complete without turning the browser request into the job lifecycle.
Finish the human workflow
Stages, voting, discussion, reminders, and notifications connected model output to the actual decision process instead of stopping at generated text.
Where it landed