INDEPENDENT PRODUCT OPS RESEARCH

90 apps. A clearer path to integration feasibility.

A source-led research workflow for mapping authentication, credential access, API surface, and buildability across the apps supplied in the assignment. Open any row to inspect the current evidence and review status.

Apps in supplied list

Draft records pending review

Source checked

Records meeting evidence checks

Needs review

Unresolved source or claim checks

Audit sample

Fixed sample for independent review

App Directory

Explore the app evidence

Source checked means each logged claim has a matching saved quote and a supporting Jev decision. Fields without enough evidence remain unknown; the manual audit is tracked separately.

Showing 90 of 90 apps

#	Platform & Domain	Category	Auth Protocols	API Breadth	Existing MCP	Buildability Verdict

Working snapshot

Patterns to verify

These counts describe current draft classifications. Source checks and the independent sample audit determine which findings can be relied on.

Category Matrix

Draft view by category

Category	Total Apps	Buildable Now	Conditional	Outreach Needed	Unknown	Self-Serve %	Dominant Auth

Verification & Quality Control

Independent accuracy audit

Pending independent review

Automated quality check Calculating saved-source quote matches...

Exact quote matching measures evidence traceability. It does not measure answer accuracy or replace the independent sample audit.

No accuracy score is available. The previous comparison used a synthetic baseline and static expected values, so it was removed rather than presented as a human audit.

The next audit should verify the same four fields for a fixed sample against official documentation, record the checked value and source URL for each field, and compare those checks with captured first-pass and final-pass outputs. Unverifiable fields must remain outside the accuracy denominator.

Sample IDs will be frozen before manual verification.

System Architecture

Multi-Stage Research & Verification Pipeline

STAGE 01 • RETRIEVAL

Source Discovery & Firecrawl Fetch

Uses a frozen list of documentation leads plus the supplied website hint. Up to two readable pages inform each record. Successful local snapshots are reused; direct HTTP runs before a budgeted Firecrawl fallback.

STAGE 02 • SYNTHESIS

OpenAI Extraction & Evidence Quoting

OpenAI proposes structured claims from bounded excerpts of saved pages. Each evidence item names its exact source URL; unmatched quotes remain unresolved rather than being presented as verified.

STAGE 03 • DECISIONS

TypeSafe Jev Atomic Verification

When an OpenRouter key is configured, the pipeline asks TypeSafe Jev to check individual source-backed claims. Without a key, claim verification remains insufficient rather than being marked supported.

STAGE 04 • COMPOSITION

Code-Owned Deterministic Rules

Python code owns the business logic for calculating api_breadth and final buildability status based on confirmed credential access conditions, eliminating arbitrary LLM guessing.

STAGE 05 • VALIDATION

Deterministic Schema & Evidence Audit

Checks schema shape, basic contradictions, evidence URLs, and exact quote matches against cached snapshots. Any missing source or quote mismatch is reported as an issue and fails validation.

STAGE 06 • PUBLISHING

Static Compilation & Web Deployment

Builds summary JSON, app records JSON/CSV, and a local JavaScript bundle from the records. The public page serves these static files without exposing or using API keys.

Runnable CLI Triggers & Execution Commands

The CLI checkpoints each app. Request limits bound paid calls, and the independent sample audit remains separate from automated source checking.

# 1. Investigate a single application by ID (e.g. #81 Stripe)

$ python -m src.research --app-id 81

# 2. Run or resume the complete 90-app automated research agent

$ python -m src.research --all --resume --max-firecrawl-calls 20 --max-model-calls 90 --max-jev-calls 360

# 3. Deterministically validate schemas, quotes, and rule consistency

$ python -m src.validate

# 4. Score only after the human audit manifest has field values and source URLs

$ python -m src.audit

# 5. Compile static analytics and regenerate frontend datasets

$ python -m src.analyze

Transparency & Disclosures

Scope Discrepancy & System Limitations

Assignment Scope Discrepancy Disclosure

The PDF assignment prompt specifies: “You will evaluate 100 apps across 10 categories”. However, the itemized table provided in the take-home document lists only 90 applications across 9 categories, terminating abruptly at #90 PitchBook in Category 9 (Finance and Fintech).

The repository contains all 90 supplied input rows. Source-backed findings are marked per row; remaining unknowns and review items are visible in the explorer. If the missing ten entries arrive, they can be added to the input list and researched with the same pipeline.

Research Agent Limitations & Boundary Conditions

Anti-Bot & Cloudflare Challenges: Certain developer portals deploy JS challenges that block direct HTTP scraping. Firecrawl can provide an alternative source; browser review remains a manual follow-up.
Login-Walled Developer Consoles: Platforms like Gladly and PitchBook conceal developer token creation behind enterprise subscription logins; public marketing docs were utilized for evidence.
Current verification limit: Checked-in classifications are provisional. A claim remains unknown or needs review when the available source does not support it; the site does not currently report an accuracy score.