Technical Reference

Back to all tools

Every tool in the workspace is listed below with what it does, which AI model powers it, where your data is processed, and how each tool’s AI behaviour is guided.

Last updated: 2026-07-26 03:06 UTC

Table of Contents

Transcribe

Category: pipeline Type: Step-by-step workflow URL: /transcribe

Description

Upload an audio or video recording (up to 2 hours) and get a text transcript with speaker labels.

Model & Hosting

  • Model: Gladia Solaria-3 transcription with pyannote speaker diarization
  • Hosting: On-server (Canadian VPS) + Gladia API (France)

Data Handling

  • Your recording is sent to Gladia (France-hosted) for transcription and speaker detection
  • Gladia deletes uploads and transcripts according to the account retention setting; the shortest standard setting is 0 days, which deletes data within 24 hours
  • All files are automatically deleted from our server after 1 hour

System Prompt

(No system prompt ? direct file upload to Gladia transcription API with diarization enabled. Runtime transcription model defaults to solaria-3 via GLADIA_TRANSCRIPTION_MODEL. Speaker detection uses pyannote audio analysis, not an LLM.)

Technical Notes

Audio conversion via ffmpeg on-server. Transcription defaults to Gladia Solaria-3 via GLADIA_TRANSCRIPTION_MODEL; set GLADIA_TRANSCRIPTION_MODEL=solaria-1 to use the broader multilingual/code-switching model. Speaker diarization uses Gladia/pyannote audio analysis, not LLM inference. This integration currently uses Gladia file upload plus result polling, so it is not Gladia’s Zero Data Retention callback flow; Gladia deletes uploaded audio and transcripts according to the account retention setting. Gladia documents the shortest standard retention setting as 0 days, which deletes data within 24 hours. Files on our server auto-delete after session expiry (1 hour). Supports recordings up to ~2 hours.

Last verified: 2026-06-11

Remove Personal Information

Category: pipeline Type: Step-by-step workflow URL: /clean

Description

Upload transcripts, survey responses, or other text files. The tool flags personal information such as phone numbers, emails, names, job titles, and combinations that could identify someone. You review and choose what to remove.

Model & Hosting

  • Model: deepseek/deepseek-v4-pro
  • Hosting: On-server (Canadian VPS) + TensorX EU API for AI review

Data Handling

  • Step 1 → Pattern scan: on-server regex catches phone numbers, emails, postal codes, SINs, URLs, social media handles, and long ID numbers. Your data doesn’t leave our server for this step.
  • Step 2 → AI review: your text is sent over HTTPS directly to the TensorX API for review by DeepSeek V4 Pro, an open-weight model run on TensorX infrastructure. LogicalOutcomes is a paying customer of TensorX, and processing is governed by TensorX’s standard Data Processing Agreement, publicly available on their website. Under that agreement, personal data is not processed outside the EEA or UK without prior written consent, TensorX lists no non-EU model provider among its sub-processors, and prompt content and results are processed in ephemeral enclaves and are not retained after processing. (TensorX Data Processing Agreement)
  • This tool should not be used for special-category data (health, racial or ethnic origin, religious beliefs, sexual orientation), as TensorX’s Data Processing Agreement requires a separate agreement for that data, which LogicalOutcomes has not put in place.
  • Uploaded files and session data are deleted from our server after 1 hour (standard sessions) or after 3 hours (batches over ten files). A cleanup process also runs at startup and every thirty minutes.
  • Applicable law: the TensorX agreement is governed by EU and Irish data protection law. It supports LogicalOutcomes’ accountability as a processor but makes no commitments under Canadian legislation. Obligations under PIPEDA, and under PHIPA or Quebec’s Law 25 where they apply, remain with LogicalOutcomes and the uploading organization.
  • Always review cleaned files before sharing. Automated de-identification is a strong first pass, not a replacement for human review.

How It Works

Built-in scanning uses pattern-matching (regex) to find and flag common personal information like phone numbers, email addresses, social insurance numbers, postal codes, URLs, social media handles, speaker labels, and long ID numbers ? this runs entirely on-server with no AI involved. AI review adds a second pass using a privacy-focused AI model trained to catch identifiers that patterns miss: names, initials, job titles that narrow identification, location references, anecdotes about specific events, small-group relationships, and combinations of details that could identify someone in a small organisation. The current AI contract is findings-only: the model identifies exact PII strings, then server-side code applies consistent markers ([NAME-1], [LOCATION-1], etc.) and presents proposed replacements for review before final output is written. Users can preserve organisation or location context when it is needed for analysis.

Full system prompt available on request for authorized reviewers.

Technical Notes

Pattern-based PII detection uses regex only (no external calls). AI review uses DeepSeek V4 Pro via TensorX (EU infrastructure), governed by TensorX’s Data Processing Agreement, publicly available on their website — no processing outside the EEA/UK without written consent, no non-EU sub-processors, and no retention after processing in TensorX’s ephemeral enclaves. Runtime model ID: deepseek/deepseek-v4-pro. The LLM returns findings only; server-side code applies replacements, preserves selected organisation/location context where requested, detects cross-chunk/cross-batch missed names, and pauses for user review before final cleaned output is written. Special-category GDPR data (health, ethnicity, religion, sexual orientation) is out of scope pending a supplementary agreement with TensorX. The TensorX agreement is governed by EU and Irish law and does not substitute for LogicalOutcomes’ or the uploading organization’s obligations under PIPEDA, PHIPA, or Quebec’s Law 25.

Last verified: 2026-07-08

Qualitative Analysis

Category: pipeline Type: Step-by-step workflow URL: /qual

Description

Upload transcripts or survey responses. The tool identifies patterns and suggests themes for your key evaluation questions. Provides a report with themes, supporting quotes and questions for discussion.

Model & Hosting

  • Model: google/gemini-3.1-pro-preview
  • Hosting: On-server (Canadian VPS) + OpenRouter API

Data Handling

  • Your text is sent to Google Gemini (an AI service) to help find themes and draft findings. By this stage, you should have already used the Transcribe and Data Cleaning tools to remove names and other personal information from your files.
  • You review and approve every step before you download your Excel workbook.
  • All project data is automatically deleted from our server after 7 days of inactivity ? save your exports before then.
  • You can delete your project at any time from the Results page.

How It Works

The qualitative analysis workflow uses AI to produce evaluation-question findings from cleaned qualitative data:

  • Analysis framework: you choose or enter evaluation questions, objectives, data source type, and optional context such as program stage, information sources, and reference documents. Reference documents are used only as background; they are not coded, quoted, or counted as participant evidence.
  • Data-source-aware coding: AI adjusts instructions for interviews, focus groups, meeting notes, open-ended survey responses, public comments, and repeated interviews with the same person over time.
  • Speaker roles: when speaker roles are assigned, the prompt tells the model which labelled speakers are respondent voices and which are interviewer/context voices.
  • Findings by question: AI produces findings only where the data provides evidence. Each finding includes what’s working, what needs attention, specific excerpts, a data-grounded discussion prompt, evaluation-objective tags, themes, and a coverage list of every participant or source expressing each theme.
  • Batch reconciliation and synthesis: for larger projects, batches are coded independently, then reconciled into a unified set of findings before per-question narratives are written for the Excel workbook.

All AI outputs are presented as drafts for human review. Source locations for excerpts are checked by server-side matching and validation rather than relying only on model-reported positions.

Full system prompt available on request for authorized reviewers.

Technical Notes

Coding, batch reconciliation, per-question synthesis, and final summary generation use OpenRouter API with Gemini 3.1 Pro Preview (google/gemini-3.1-pro-preview). Findings are organized by evaluation question and exported to an Excel workbook for staff review. Reference documents are background only and are not coded, quoted, or counted as participant evidence. Source locations for excerpts are determined by post-hoc matching and validation, not only by LLM-reported positions.

Last verified: 2026-06-18

Evaluation Planner

Category: more-tools Type: Step-by-step workflow URL: https://evalplanner.logicaloutcomes.net/

Description

Describe your program to get a complete draft evaluation plan with a logic model, evaluation questions, indicators, and steps for implementation.

Model & Hosting

  • Model: OpenRouter (configurable per step)
  • Hosting: Railway — Toronto edge/YYZ (project: LO-EvalPlanner202510) ? evalplanner.logicaloutcomes.net

Data Handling

  • Program details are processed by AI to generate your evaluation plan (data passes through OpenRouter, a US-based AI routing service)
  • Your website may be read automatically to gather program context
  • Reports can be emailed to you or downloaded as HTML

How It Works

Enter your program details and website. AI researches your organisation, asks clarifying questions, and generates a complete evaluation plan with logic model, evaluation questions, indicators, and next steps. Reports can be emailed or downloaded as HTML.

Full system prompt available on request for authorized reviewers.

Technical Notes

React/TypeScript SPA with Express backend. AI prompts are managed in a PostgreSQL database. Background job queue handles long-running AI calls. Reports rendered as responsive HTML with logic model diagrams.

Last verified: 2026-04-03