Soji
Soji
Soji
Discover
Discover
Discover
Key Deliverables:
Key Deliverables:
→
→
→
Web application;
Web application;
→
→
→
Brand identity & brand book;
Brand identity & brand book;
→
→
→
Marketing website;
Marketing website;
→
→
→
Design system.
Design system.
Timeline:
2024 → Ongoing
Timeline:
2024 → Ongoing
Role:
Lead Product Designer
Role:
Lead Product Designer
Engagement:
End-to-end product & brand ownership
Engagement:
End-to-end product & brand ownership
Soji is a one-stop AI platform built for aviation maintenance — turning the industry's most regulated, document-heavy workflows into structured, agent-driven intelligence that maximises aircraft uptime.
Soji is a one-stop AI platform built for aviation maintenance — turning the industry's most regulated, document-heavy workflows into structured, agent-driven intelligence that maximises aircraft uptime.
Aviation maintenance operates under EASA and FAA oversight, with technicians cross-referencing thousands of pages of manuals, service bulletins, and airworthiness directives for every task. Generic AI tools hallucinate or strip the regulatory context that makes data defensible. Soji needed an interface delivering the speed of a modern AI assistant with the audit trail the industry demands by law.
Aviation maintenance operates under EASA and FAA oversight, with technicians cross-referencing thousands of pages of manuals, service bulletins, and airworthiness directives for every task. Generic AI tools hallucinate or strip the regulatory context that makes data defensible. Soji needed an interface delivering the speed of a modern AI assistant with the audit trail the industry demands by law.
<60s
<60s
Time to first verified answer
Time to first verified answer
100%
100%
Source-traced AI responses
Source-traced AI responses
6
6
Integrated modules
Integrated modules
Define
Define
Define
Soji enters a vertical that has resisted modern software for decades. Legacy MRO platforms (AMOS, TRAX, Ramco) automate record-keeping but do nothing for the actual cognitive work: finding the right procedure, in the right revision, for the right aircraft. General-purpose AI is worse — it hallucinates part numbers and strips source traceability. In a vertical where a misapplied procedure can ground a fleet, probably correct isn't a product.
Soji enters a vertical that has resisted modern software for decades. Legacy MRO platforms (AMOS, TRAX, Ramco) automate record-keeping but do nothing for the actual cognitive work: finding the right procedure, in the right revision, for the right aircraft. General-purpose AI is worse — it hallucinates part numbers and strips source traceability. In a vertical where a misapplied procedure can ground a fleet, probably correct isn't a product.
18 expert interviews
18 expert interviews
We spoke with maintenance engineers, compliance officers, lease return managers, and MRO operations leads across Hamburg, Singapore, and Dubai. We expected the core problem to be speed. It wasn't. Engineers didn't want a faster search — they wanted an answer they could sign off on: exact manual, exact revision, exact paragraph, chain intact. A fast answer without provenance was worse than no answer, because it created the illusion of work done.
We spoke with maintenance engineers, compliance officers, lease return managers, and MRO operations leads across Hamburg, Singapore, and Dubai. We expected the core problem to be speed. It wasn't. Engineers didn't want a faster search — they wanted an answer they could sign off on: exact manual, exact revision, exact paragraph, chain intact. A fast answer without provenance was worse than no answer, because it created the illusion of work done.
This led us to ask: "How might we let AI accelerate aviation maintenance decisions without compromising the audit trail that keeps aircraft legally airworthy?"
This led us to ask: "How might we let AI accelerate aviation maintenance decisions without compromising the audit trail that keeps aircraft legally airworthy?"
Of six planned modules, Troubleshooting Support was prioritised first — highest-frequency, highest-pressure use case, clearest measurable outcome, fastest path to validating the thesis on the hangar floor.
Of six planned modules, Troubleshooting Support was prioritised first — highest-frequency, highest-pressure use case, clearest measurable outcome, fastest path to validating the thesis on the hangar floor.
Personas
Personas
Anja, the Line Maintenance Engineer (34, Senior Aircraft Technician):
Anja, the Line Maintenance Engineer (34, Senior Aircraft Technician):
→
→
Works under AOG pressure, where every minute of downtime costs the operator thousands;
Works under AOG pressure, where every minute of downtime costs the operator thousands;
→
→
Cross-references multiple manuals on a tablet next to the aircraft, often in noisy hangar conditions;
Cross-references multiple manuals on a tablet next to the aircraft, often in noisy hangar conditions;
→
→
Needs answers fast — but cannot sign off on a task card without a verifiable source citation.
Needs answers fast — but cannot sign off on a task card without a verifiable source citation.
Rajesh, the Lease Return Manager (39, Asset Manager):
Rajesh, the Lease Return Manager (39, Asset Manager):
→
→
→
Manages the months-long redelivery process, reconciling discrepancies across thousands of records before handing the aircraft back to its lessor;
Manages the months-long redelivery process, reconciling discrepancies across thousands of records before handing the aircraft back to its lessor;
→
→
→
Coordinates across maintenance, finance, legal, and the lessor — each with different definitions of compliant;
Coordinates across maintenance, finance, legal, and the lessor — each with different definitions of compliant;
→
→
→
Needs structured workflows that turn unstructured documentation into a defensible redelivery package.
Needs structured workflows that turn unstructured documentation into a defensible redelivery package.
Maria, the MRO Operations Lead (45, Director of Maintenance Operations):
Maria, the MRO Operations Lead (45, Director of Maintenance Operations):
→
→
→
Quotes maintenance work to airlines and operators, balancing scope, cost, and turnaround time;
Quotes maintenance work to airlines and operators, balancing scope, cost, and turnaround time;
→
→
→
Currently rebuilds proposals from scratch every cycle, losing institutional knowledge between bids;
Currently rebuilds proposals from scratch every cycle, losing institutional knowledge between bids;
→
→
→
Needs reusable templates that surface historical pricing, parts availability, and capacity in one place.
Needs reusable templates that surface historical pricing, parts availability, and capacity in one place.
Problem Statement
Problem Statement
Document Overload
Document Overload
Document Overload
Manuals and directives run into thousands of pages per aircraft type, scattered across PDFs and portals.
Manuals and directives run into thousands of pages per aircraft type, scattered across PDFs and portals.
Hallucination Risk
Hallucination Risk
Hallucination Risk
Generic AI paraphrased regulations and dropped citations. Probably correct is a liability, not a feature.
Generic AI paraphrased regulations and dropped citations. Probably correct is a liability, not a feature.
Broken Audit Trails
Broken Audit Trails
Broken Audit Trails
Compliance officers reconstructed decisions weeks after the fact. If the chain couldn't be traced, auditors treated it as if it hadn't happened.
Compliance officers reconstructed decisions weeks after the fact. If the chain couldn't be traced, auditors treated it as if it hadn't happened.
Knowledge Loss Between Cycles
Knowledge Loss Between Cycles
Knowledge Loss Between Cycles
Redelivery packages and proposals were rebuilt from scratch each time, living in senior engineers' heads.
Redelivery packages and proposals were rebuilt from scratch each time, living in senior engineers' heads.
How might we?
How might we?
→
→
→
HMW give engineers a single answer for any maintenance question, citing the exact manual, revision, and paragraph?
HMW give engineers a single answer for any maintenance question, citing the exact manual, revision, and paragraph?
→
→
→
HMW make every AI-generated decision defensible in front of an EASA or FAA auditor by default?
HMW make every AI-generated decision defensible in front of an EASA or FAA auditor by default?
→
→
→
HMW turn months of unstructured aircraft documentation into a structured redelivery package at end-of-lease?
HMW turn months of unstructured aircraft documentation into a structured redelivery package at end-of-lease?
→
→
→
HMW let MRO operations teams reuse pricing, scope, and capacity knowledge across maintenance proposals instead of rebuilding each one?
HMW let MRO operations teams reuse pricing, scope, and capacity knowledge across maintenance proposals instead of rebuilding each one?
→
→
→
HMW design a platform that earns the trust of a hangar engineer and a compliance auditor with the same interface?
HMW design a platform that earns the trust of a hangar engineer and a compliance auditor with the same interface?
Key metrics
Key metrics
To track success, we monitored:
To track success, we monitored:
→
→
→
Time to verified answer — Median time from query to a cited, source-traced response on the hangar floor;
Time to verified answer — Median time from query to a cited, source-traced response on the hangar floor;
→
→
→
Source traceability rate — Share of AI responses returning a verifiable manual, revision, and paragraph citation;
Source traceability rate — Share of AI responses returning a verifiable manual, revision, and paragraph citation;
→
→
→
Audit reconstruction time — How long it takes a compliance officer to rebuild the decision chain for any past task;
Audit reconstruction time — How long it takes a compliance officer to rebuild the decision chain for any past task;
→
→
→
Module reuse rate — How often a redelivery package, proposal, or troubleshooting playbook is reused vs. rebuilt;
Module reuse rate — How often a redelivery package, proposal, or troubleshooting playbook is reused vs. rebuilt;
→
→
→
AOG resolution time — Median time from a grounded aircraft to a documented, signed-off return-to-service decision.
AOG resolution time — Median time from a grounded aircraft to a documented, signed-off return-to-service decision.
Iterate
Iterate
Iterate
Our design process balanced technically demanding concepts — semantic search across regulated documentation, agent-driven workflows — with an experience approachable for engineers under AOG pressure on a hangar floor. The challenge wasn't designing for non-technical users — it was designing for technical users who couldn't afford to be wrong.
Our design process balanced technically demanding concepts — semantic search across regulated documentation, agent-driven workflows — with an experience approachable for engineers under AOG pressure on a hangar floor. The challenge wasn't designing for non-technical users — it was designing for technical users who couldn't afford to be wrong.
First Iteration:
First Iteration:
→
→
→
Started with an open chat interface — free-form questions, synthesised answers from the full document corpus;
Started with an open chat interface — free-form questions, synthesised answers from the full document corpus;
→
→
→
Treated source citations as a footer artifact, visually de-prioritised against the AI response;
Treated source citations as a footer artifact, visually de-prioritised against the AI response;
→
→
→
Validated that AI could synthesise faster than manual search — but exposed friction the moment engineers tried to act on it.
Validated that AI could synthesise faster than manual search — but exposed friction the moment engineers tried to act on it.
Second Iteration:
Second Iteration:
Testing revealed a consistent pattern: trust came from evidence shown first, answer shown second.
Testing revealed a consistent pattern: trust came from evidence shown first, answer shown second.
→
→
→
Inverted the answer hierarchy — cited manual, revision, and paragraph render first, synthesised answer as supporting context;
Inverted the answer hierarchy — cited manual, revision, and paragraph render first, synthesised answer as supporting context;
→
→
→
Replaced open chat with a structured task interface — aircraft type, ATA chapter, task context upfront;
Replaced open chat with a structured task interface — aircraft type, ATA chapter, task context upfront;
→
→
→
Introduced a persistent audit panel showing which documents were consulted and where the agent deferred to a human;
Introduced a persistent audit panel showing which documents were consulted and where the agent deferred to a human;
→
→
→
Restructured navigation around the maintenance lifecycle — Troubleshooting primary, other modules connected through the same knowledge base.
Restructured navigation around the maintenance lifecycle — Troubleshooting primary, other modules connected through the same knowledge base.
These iterations reframed the product from AI that answers aviation questions to a system that produces decisions a regulator can defend — moving Soji out of the AI-assistant category and into the operational stack alongside AMOS, TRAX, and Ramco.
These iterations reframed the product from AI that answers aviation questions to a system that produces decisions a regulator can defend — moving Soji out of the AI-assistant category and into the operational stack alongside AMOS, TRAX, and Ramco.
Design
Design
Design
Soji didn't yet have a brand, a design system, or a product surface — the three had to be built as one continuous problem, not in sequence.
Soji didn't yet have a brand, a design system, or a product surface — the three had to be built as one continuous problem, not in sequence.
Brand Identity & Visual System:
Brand Identity & Visual System:
→
→
→
A confident, industrial visual language — Soji needed to read as aerospace-grade infrastructure, not another AI startup;
A confident, industrial visual language — Soji needed to read as aerospace-grade infrastructure, not another AI startup;
→
→
→
A near-black foundation with a cool teal-cyan accent applied to verified information and AI-generated outputs;
A near-black foundation with a cool teal-cyan accent applied to verified information and AI-generated outputs;
→
→
→
A full brand book covering logo, type, colour, motion, and voice — a single source of truth across product, marketing, sales, and recruitment.
A full brand book covering logo, type, colour, motion, and voice — a single source of truth across product, marketing, sales, and recruitment.
Design System:
Design System:
→
→
→
One system spans product and marketing — every component travels from the landing hero to the deepest audit panel;
One system spans product and marketing — every component travels from the landing hero to the deepest audit panel;
→
→
→
Component primitives are workflow-specific — Source Citation, Verified Answer, Audit Trail Entry, Task Card are first-class, not generic patterns;
Component primitives are workflow-specific — Source Citation, Verified Answer, Audit Trail Entry, Task Card are first-class, not generic patterns;
→
→
→
Density is a first-class dimension — comfortable, compact, and hangar-floor variants of every component.
Density is a first-class dimension — comfortable, compact, and hangar-floor variants of every component.
User Experience & Information Architecture:
User Experience & Information Architecture:
→
→
→
Every AI response leads with the evidence — citation first, synthesised answer as supporting context;
Every AI response leads with the evidence — citation first, synthesised answer as supporting context;
→
→
→
A persistent audit trail shows which documents the AI consulted and where it deferred to a human — turning "the AI said so" into a defensible chain that holds up under EASA or FAA review.
A persistent audit trail shows which documents the AI consulted and where it deferred to a human — turning "the AI said so" into a defensible chain that holds up under EASA or FAA review.
Adoption & Distribution:
Adoption & Distribution:
→
→
→
Onboarding centers on one milestone — a maintenance question answered with full source traceability inside the first session;
Onboarding centers on one milestone — a maintenance question answered with full source traceability inside the first session;
→
→
→
The design system extends to sales and partner materials, so every touchpoint reinforces the same thesis;
The design system extends to sales and partner materials, so every touchpoint reinforces the same thesis;
→
→
→
In a vertical that evaluates vendors against decade-old incumbents, a coherent visual system communicates seriousness before a single feature is demoed.
In a vertical that evaluates vendors against decade-old incumbents, a coherent visual system communicates seriousness before a single feature is demoed.
Measure & Test
Measure & Test
Measure & Test
Data Analysis & Testing Approach:
Data Analysis & Testing Approach:
→
→
→
Instrumented every query, citation render, and audit panel open via PostHog;
Instrumented every query, citation render, and audit panel open via PostHog;
→
→
→
Citation-first responses were verified and acted on 2.8x more often than answer-first responses;
Citation-first responses were verified and acted on 2.8x more often than answer-first responses;
→
→
→
Engineers tolerated a slower response with a verifiable source, but abandoned the workflow after a single uncited reference.
Engineers tolerated a slower response with a verifiable source, but abandoned the workflow after a single uncited reference.
Enhancements Following EASA Alignment:
Enhancements Following EASA Alignment:
→
→
→
Per-aircraft access controls, data residency signalling, and a full audit log surfaced directly in the workspace UI;
Per-aircraft access controls, data residency signalling, and a full audit log surfaced directly in the workspace UI;
→
→
→
Document revision tracking visible at every AI response — engineers see whether the cited manual is the active revision;
Document revision tracking visible at every AI response — engineers see whether the cited manual is the active revision;
→
→
→
Compliance questions that once meant weeks of back-and-forth were increasingly answered by the interface itself.
Compliance questions that once meant weeks of back-and-forth were increasingly answered by the interface itself.
User Research & Iterative Refinements:
User Research & Iterative Refinements:
→
→
→
Early prototypes leaned consumer-AI. Interviews revealed the opposite preference: "If it talks like a chatbot, I treat it like a chatbot," one Head of Continuing Airworthiness said;
Early prototypes leaned consumer-AI. Interviews revealed the opposite preference: "If it talks like a chatbot, I treat it like a chatbot," one Head of Continuing Airworthiness said;
→
→
→
Microcopy was rewritten in the voice of a senior engineer — error states explain which document failed and why;
Microcopy was rewritten in the voice of a senior engineer — error states explain which document failed and why;
→
→
→
Restructuring the task interface increased first-query completion by 28% among engineers, with no drop among non-technical staff.
Restructuring the task interface increased first-query completion by 28% among engineers, with no drop among non-technical staff.
Integrations & Credibility:
Integrations & Credibility:
→
→
→
Named integrations with AMOS, TRAX, and Ramco made clear from the first screen that Soji extends the existing stack;
Named integrations with AMOS, TRAX, and Ramco made clear from the first screen that Soji extends the existing stack;
→
→
→
Regulatory alignment (EASA Part-145, Part-M, ISO 27001 in progress) surfaced as live signals in marketing and sales materials;
Regulatory alignment (EASA Part-145, Part-M, ISO 27001 in progress) surfaced as live signals in marketing and sales materials;
→
→
→
The brand book itself became a sales artifact — a one-page extract shown before any product demo.
The brand book itself became a sales artifact — a one-page extract shown before any product demo.
The Measure & Test phase shifted Soji from a faster way to search documents to the system operators trust to defend a decision. The metrics that mattered: citation verification rate, first-answer accuracy, and audit reconstruction time.
The Measure & Test phase shifted Soji from a faster way to search documents to the system operators trust to defend a decision. The metrics that mattered: citation verification rate, first-answer accuracy, and audit reconstruction time.
Impact
Impact
Impact
Soji was built from zero. By the end of the engagement, the platform had a working AI module on the hangar floor, a visual identity operators recognised as serious infrastructure, and a design system extensible across five more modules.
Soji was built from zero. By the end of the engagement, the platform had a working AI module on the hangar floor, a visual identity operators recognised as serious infrastructure, and a design system extensible across five more modules.
The number that mattered most
The number that mattered most
Every AI response ships with a verifiable citation — manual, revision, paragraph, every time. 100% source traceabilityisn't a design metric in most products. In aviation maintenance, it is the product. Designing around citation-first rendering turned this from an aspiration into a default.
Every AI response ships with a verifiable citation — manual, revision, paragraph, every time. 100% source traceabilityisn't a design metric in most products. In aviation maintenance, it is the product. Designing around citation-first rendering turned this from an aspiration into a default.
The behavioural shift
The behavioural shift
Engineers used to re-read every response twice — once to understand, once to verify. After inverting the hierarchy, that second pass disappeared. "I read the source first now, then the answer if I need it," one engineer said during testing.
Engineers used to re-read every response twice — once to understand, once to verify. After inverting the hierarchy, that second pass disappeared. "I read the source first now, then the answer if I need it," one engineer said during testing.
The signal from the industry
The signal from the industry
Pilot operators completed procurement and compliance review in the same week of first contact — a cycle that typically runs four to six weeks in aviation software. Multiple prospects cited visual coherence as the reason they took the team seriously.
Pilot operators completed procurement and compliance review in the same week of first contact — a cycle that typically runs four to six weeks in aviation software. Multiple prospects cited visual coherence as the reason they took the team seriously.
The foundation, not the finish
The foundation, not the finish
Five modules beyond Troubleshooting are scoped to ship on the same component primitives and audit-first model — a compounding advantage that won't fully show up in a metric until 2026.
Five modules beyond Troubleshooting are scoped to ship on the same component primitives and audit-first model — a compounding advantage that won't fully show up in a metric until 2026.
Aviation maintenance has decades of competent, indistinguishable enterprise tools operators tolerate. The opportunity for Soji was to be the first platform in the category that earned trust through how it looked, how it answered, and how it could be defended — in the same gesture.
Aviation maintenance has decades of competent, indistinguishable enterprise tools operators tolerate. The opportunity for Soji was to be the first platform in the category that earned trust through how it looked, how it answered, and how it could be defended — in the same gesture.
Copyright © 2026 Monochromatic Ltd. All rights reserved.
Workhub, 77 Lower Camden Street, Dublin, D02 XE80, Ireland
Copyright © 2026 Monochromatic Ltd. All rights reserved.
Workhub, 77 Lower Camden Street, Dublin, D02 XE80, Ireland
Copyright © 2026 Monochromatic Ltd. All rights reserved.
Workhub, 77 Lower Camden Street, Dublin, D02 XE80, Ireland