DELPHI Ai is AI search for aviation manuals: flight, cabin, maintenance and dispatch staff ask a question and get an answer from your own manuals, cited to the document, section and page. It searches only the revisions an administrator has published, and an automated check labels whether the cited passages support each manual answer. When the evidence is weak, DELPHI Ai shows what it found instead of guessing.
Email me the brochure, DELPHI Ai
DELPHI Ai is FlightAtom's aviation knowledge AI: cited search and question answering over an operator's own manuals. It answers from the revisions an administrator has reviewed and published, ties each answer to the document, section and page it came from, and declines to answer when the evidence is weak or it cannot confirm that a procedure is current. Flight, cabin, maintenance and dispatch staff, instructors and safety teams each search only the documents their role allows.
Writing the answer is one of these steps. The others decide which manual text DELPHI Ai may draw on, whether it answers at all, which label the answer carries, and how a reader checks it.
DELPHI Ai screens with demonstration manuals and questions. The answers shown were scripted for the demonstration rather than generated live. Open a screenshot to inspect the citation and the source behind it.
A question about a 120 Wh power bank gets an answer labelled Grounded from manuals, with an [S#] marker on every statement. Both cited sections are listed with page and relevance score, and the first is open at its passage and PDF link. Synthetic demonstration data.
A single-runway contingency fuel question is answered from Notice to Staff 26-014 alone, with a source note that the notice replaces OM-A 8.3.3. Under the Amendment in effect banner, the old OM-A paragraph stays listed but is marked Superseded. Synthetic demonstration data.
A broad question gets no generated answer: DELPHI Ai marks it Needs clarification with low confidence, asks for the exact procedure name or keyword, and lists the passages it found across the manuals. Synthetic demonstration data.
Revision 4 of a demonstration operations manual is In review and unavailable to users while an administrator checks the original cover, records a reason and confirms the revision before approving it. Synthetic demonstration data.
Asked with Source set to Auto, a flight-data question is routed to the flight database and labelled Grounded from flight DB. The route figures are in table view, with the read-only SQL behind them open under Source: flight database query. Synthetic demonstration data.
Query Insights groups the questions people ask into topics with a suggested action for each. The filtered list shows de-icing holdover times as a gap the manuals do not cover. Synthetic demonstration data.
Each capability keeps an answer tied to its source, and the source under the operator's control.
A manual answer is labelled Grounded from manuals only when every substantive sentence is cited and an automated check confirms the cited passage supports it; otherwise it is marked as a partial answer to verify against the sources, or as one DELPHI Ai could not verify. Each [S#] citation shows the document, section and page, flags passages taken from tables or figures, and puts the passage and the original file a click away.
When the evidence is weak, DELPHI Ai writes no answer: it shows the passages it found and asks for the exact procedure name or keyword, suggesting the aircraft type or phase of flight if relevant. It also holds back when it cannot confirm a procedure is the current version, and says when a search service is down instead of reporting that nothing was found. Administrators can make cited passages without generated text the default.
Uploads enter as drafts. A revision becomes searchable only after an administrator reviews the exact original file, confirms its revision number and effective date, and publishes it on or after that date; approval is tied to that file, so a changed original has to be approved again. A replacement marks the earlier revision superseded, and every step is recorded with a reason.
For procedure questions, a relevant Notice to Staff is weighted above base manual text, and the answer model is instructed that a notice takes precedence over the procedure it amends. Its citations carry an NTS badge under an Amendment in effect notice. When an amendment is linked to the procedure it replaces, the older passage is marked superseded but stays visible, so readers can see both.
Query Insights uses an AI model to group the questions people ask into training topics with a suggested action for each, weighted toward low-confidence answers, retries and negative feedback, and shows whether searches found good, weak or no manual coverage for each topic. A Training Analysis dashboard counts questions by subject, manual and week, and shows the cited-answer rate and which subjects drew unanswered or weak answers. Administrators can also schedule cited email reports and Daily Knowledge briefings, which rotate sources and avoid repeating recent topics.
Cabin, maintenance and dispatch staff search a default set of document types for their role, such as OM-A and DGM for cabin, AMM for maintenance and MEL for maintenance and dispatch, and occurrence reports stay with safety and administrator roles. Permissions apply before retrieval, so a passage a user may not see never reaches that user's answers. Administrators refine access by document, collection and profile.
Authorized analysts can query the manuals or connected flight, training and hiring databases from the same box, or combine manual guidance with flight data in one answer. DELPHI Ai routes each question to the right source, answers data questions with read-only queries, and applies the same citation check to combined answers. Flight-data results can be charted, switched to a table and traced to the query behind them.
Administrators check selected ISARPs against the operator's manuals in batches and export the results. An AI model classes each as documented, partially documented or not found, noting where the manuals cover it and the gaps to follow up. Standard text is kept apart from operator evidence, and any documented or partially documented result without cited operator evidence is rejected automatically, so requirement text alone never counts as evidence.
Learning sessions teach from the operator's own manuals, explaining the reasoning behind procedures and naming what the material does not cover, and trainee accounts can be limited to them. Focused sessions let administrators, training staff and executives summarize, explain or compare the documents they pick. Training staff can plan from MENTHOR Ai rosters, qualifications, due dates and simulator pairings, and the training calendar changes only when a request explicitly asks.
Most of DELPHI Ai's AI work happens around the answer: preparing manuals so a question phrased the way crews ask finds text phrased the way manuals are written, finding the passage that settles it, and holding the answer to that passage.
DELPHI Ai ingests PDF, DOCX, PPTX, XLSX, PNG and JPEG files, up to 50 at a time, and recovers scanned PDF pages with OCR. Detected WARNING, CAUTION and NOTE blocks, checklists and tables stay together as single passages, and an AI model adds a summary, likely questions and the aircraft, procedure type and flight phase to each one. Administrators can also send selected PDFs through vision extraction for complex tables and layouts.
Each question is expanded for aviation acronyms and synonyms and checked for aircraft and systems, then searched by meaning and by keyword together. An AI relevance check screens out passages that do not address it, and a bounded retry refines the search terms when the first results are poor. Thumbs-up and thumbs-down ratings on earlier answers nudge the cited documents and passages up or down in later rankings.
The answer model is instructed to quote limits and values exactly as written, attribute each procedure to its document and section, ask for the aircraft variant or flight phase when the answer depends on it, and never blend superseded and current values. Manual text reaches the model labelled as reference data, with an instruction never to follow directives inside it.
Agree a representative sample and a baseline from your current process. These are proposed evaluation measures, not published performance results.
Approved manuals with conflicting revisions and difficult or unanswerable questions.
Check sources, permissions, controlling versions, and unsupported answers.
Citation support; revision correctness; unsupported answers; time to verify.
DELPHI Ai declines to write an answer when the retrieved evidence is weak. It shows the passages it did find, so you can inspect them, and asks for the exact procedure name or keyword, suggesting the aircraft type or phase of flight if relevant. It also declines when it cannot confirm that a retrieved procedure is the current version. Administrators can go further and make cited passages without generated text the default.
Answers cite their sources with [S#] markers, and you can expand each cited passage or open the original file at the cited page. A manual answer is labelled Grounded from manuals only when every substantive sentence is cited and an automated check confirms the cited passage supports it. Otherwise it is labelled as a partial answer to verify against the sources, or as one DELPHI Ai could not verify.
Published revisions only. New uploads enter as drafts, and an administrator reviews the exact original file, confirms its revision number and effective date, and publishes it on or after that date. Publishing a replacement marks the earlier revision superseded, and published history is withdrawn rather than deleted.
Citations from Notices to Staff carry an NTS badge, and the answer shows an Amendment in effect notice. For procedure questions, a relevant notice is weighted above base manual text, and the answer model is instructed that a notice takes precedence over the procedure it amends. When an administrator links an amendment to the procedure it replaces, the older passage is marked superseded but stays visible, and the model is instructed never to present a superseded value as current or blend it with the active one.
Yes. DELPHI Ai includes a hosted MCP server with a read-only retrieval tool for authorized AI assistants. It returns cited excerpts from the published library, with document, section and page details, under the calling account's own document permissions. It does not generate an answer, so the assistant works from the source text.
Yes, for authorized analysts. DELPHI Ai routes a question to the manuals, to connected flight, training or hiring databases, or to manuals and flight data together, and answers data questions with read-only database queries. Combined answers go through the same citation check, and flight-data results can be charted, switched to a table and traced to the query behind them.
Yes, for finding and using controlled knowledge. DELPHI Ai ingests PDF, DOCX, PPTX, XLSX and image files into a revision-controlled library with effective dates, collections and role-based access, answers questions from it with citations, and shows which questions point to unclear content. FOLIO Ai is FlightAtom's separate product for authoring manuals and releasing them to readers.
No. DELPHI Ai is a standalone product with its own document library, revision control, access controls and conversations. Where an operator connects them, it also supplies cited manual retrieval and investigation drafting to TRACE Ai, uses read-only FLIGHT Ai tools for flight-operations sessions, and plans training with MENTHOR Ai data.
As your fleet, routes, rules or teams change, we adapt DELPHI Ai to match, and you approve each change before it goes live.
A focused walkthrough, with the people and decisions that matter to your team.