AI and APIs - Ensuring Platform Migration Reliability with AI

SNCF Connect & Tech is migrating its massive SaaS distribution platform to a new API-first orchestration architecture. With hundreds of millions of tickets at stake, ensuring non-regression is a mission-critical challenge. Cyrille Martraire and Thomas Nansot (Arolla) explain how they used Agentic AI and the Model Context Protocol (MCP) to automate JSON analysis and mitigate migration risks.

Key Takeaways:
• Scalable Non-Regression: How to use AI to analyze massive JSON payloads across legacy and new API formats.

• Agentic Workflows with MCP: Real-world application of the Model Context Protocol to build iterative, learning AI agents.

• Mitigating Migration Risk: Practical strategies for transitioning complex SaaS architectures without service disruption.

• Engineering Conviction: Lessons learned from technical "dead ends" and how AI is redefining complex system architecture.

------------------------------------

Conference Details:
Conference: apidays Paris 2025 | part of FOST 2025
Theme: The APIs meet AI conference: Innovation, Security, Sovereignty, Sustainability
Date: 9 - 11 December 2025 • CNIT Forest, Paris

--------------------------

Resources from apidays:
Join our upcoming conferences: https://www.apidays.global/
Read the latest API news: https://www.apiscene.io
Explore the API Landscape: https://apilandscape.apiscene.io/

Transcript

1. AI & APIs: Ensuring Platform Migration Reliability with AI Thomas Nansot - Cyrille Martraire
2. Are AI projects like any other project?
3. SUCCESS = delivering what’s planed, on-time?
4. ?
5. Let’s share our story
6. Meet Thomas
7. Meet Thomas • Director of Engineering • 150 people • API == the product
8. Distribution Engine for Mobility
9. MIGRATION project HUGE
10. @cyriux @arollafr SHIP THE CRITICAL PROJECT! my manager ✋
11. @cyriux @arollafr DO AI! my manager ✋
12. @cyriux Can’t do both…
13. ⚠ Risky migration project
14. ⚠ ❌ Only MUST-HAVE projects left in backlog Risky migration project
15. ⚠ ❌ Only MUST-HAVE projects left in backlog Risky migration project AND NOW THEY WANT EVEN MORE WITH AI!
16. I’M STUCK #stressed
17. @cyriux I need my dopamine shot!
18. LinkedIn post
19. LinkedIn post AI & Legacy Cyrille & Arolla
21. No time internally?
22. No time internally? Call in experts to do the AI stuff!
23. Meet Cyrille • Co-founder Arolla • 120 people • Software craft expert • 2 books
24. @cyriux BUY OUR BOOKS!
25. I see a potential for AI
26. Tell me more…
27. PLATFORM MIGRATION
28. every minute LEGACY API +500M€ of order per month from an existing platform
29. every minute to a new platform LEGACY API +500M€ of order per month NEW API MIGRATION
30. to a new platform LEGACY API +500M€ of order per month NEW API MIGRATION
31. to a new platform +500M€ of order per month NEW API
32. LEGACY API NEW API from legacy JSON contract Legacy JSON Contract
33. LEGACY API NEW API from legacy JSON contract Legacy JSON Contract New JSON Contract completely different
34. LEGACY API NEW API from legacy JSON contract Legacy JSON Contract New JSON Contract TBD 🚧
35. - LARGE JSON FILES - COMPLEX BUSINESS DOMAIN - MILLIONS OF TRANSACTIONS The Scale
36. MIGRATION SUCCESS ≈ CONFIDENCE IN NEW PLATFORM
37. the key: assessing trust LEGACY API NEW API ✅ trusted
38. LEGACY API NEW API 😰 trusted? the key: assessing trust
39. @AROLLAFR Am I risking my career? I'M MIGRATING! every manager
40. NEED TO PROVE RELIABILITY & FUNCTIONAL COVERAGE
41. Traditional Approach: Testing
42. NEW API End-to-End Testing #toil Tests Cases ❓ ultra laborious & repetitive
43. 😱 SCARY AMOUNT OF WORK 😫 TOIL WORK (MAKING TESTS) The Fear
44. LEGACY API NEW API from legacy JSON contract Legacy JSON Contract New JSON Contract TBD 🚧
45. 😱 SCARY AMOUNT OF WORK 😫 TOIL WORK (MAKING TESTS) 😤 TOIL WORK++ (REDOING TESTS EVERY CONTRACT CHANGE) The Fear
46. 😱 SCARY AMOUNT OF WORK 😫 TOIL WORK (MAKING TESTS) 😤 TOIL WORK++ (REDOING TESTS EVERY CONTRACT CHANGE) The Fear
47. WAIT A MINUTE…
48. LEGACY API Legacy JSON Contract NEW API New JSON Contract DIFFERENT CONTRACTS
49. LEGACY API Legacy JSON Contract NEW API New JSON Contract SAME BEHAVIOR DIFFERENT CONTRACTS
50. - Different format ✓ - Identical semantics ✓
51. LEGACY API NEW API PERFECT CANDIDATE FOR AI!
52. We’re going to use AI to make possible what was prohibitively laborious before! 👉
53. One week later…
54. We’ve done a quick prototype
55. @cyriux The initial prototype (TM)
56. LEGACY API NEW API 1/ QUERY LEGACY API (FAKE)
57. LEGACY API NEW API 1/ QUERY LEGACY API
58. LEGACY API NEW API 2/ QUERY NEW API (FAKE)
59. LEGACY API NEW API 2/ QUERY NEW API
60. 3/ ASSERT BOTH JSON HAVE SAME SEMANTICS =?=
61. IT "WORKS" =?=
62. Ok, that’s encouraging.
63. Let’s make it real!
64. From a Proof of Concept to a Project…
65. END-TO-END TESTING WITH AI
66. @cyriux
67. @cyriux The Arolla Team: Olivier +Cyrille Olivier Penhoat Senior Software Eng. (one of the finest)
68. @cyriux The Arolla Team: Olivier +Cyrille Olivier Penhoat Senior Software Eng. (one of the finest)
69. @cyriux @arollafr We’ve got a prototype,
70. @cyriux @arollafr Just plug & adapt to the real systems!
71. @cyriux Tech Challenges
72. @cyriux Knowledge puzzle distributed across many experts
73. @cyriux - Finding JSON samples - URLs of different API versions - Sample scenarios (from Postman) - Documentation (wiki, OpenAPI…)
74. WHAT TECH STACK? 😎
75. TECH STACK: A MOVING TARGET trying various tools & models…
76. Prompts & context files 👍
77. Cline just makes everything work better Agent with MCP (VsCode plugin)
78. Cline just makes everything work better Agent with MCP (VsCode plugin) (And more $$$ 😅)
79. SO MANY POSSIBILITIES 😎
80. Huge solution space EASY TO JUST GET LOST
81. We explored many options, lots of dead ends. (From now on: focus on what works) (…)
82. LARGE JSON FILES 😎
83. Large files (200 KB - 480KB) - Don't fit well in context - Too many hallucinations (Attention Dilution)
84. First ideas Slice into chunks & recombine results
85. First ideas Slice into chunks & recombine results 🤮
86. a BETTER approach
87. MCP!
88. 👉 JSONpath Query via MCP 🤖 Large JSON LLM in a agentic loop - destination - date - price
89. 👉 JSONpath Query via MCP 🤖 Large JSON LLM in a agentic loop MCP jsonpath - destination - date - price
90. 👉 JSONpath Query via MCP 🤖 Large JSON LLM in a agentic loop MCP jsonpath - destination - date - price
91. 👉 JSONpath Query via MCP 🤖 Large JSON LLM in a agentic loop - destination - date - price
92. 👉 JSONpath Query via MCP 🤩 No need to put everything in context! JSON is like a database for the agent loop.
93. Trying a good MCP server 🤩 jsontools MCP server
94. DATA DISCREPANCIES 😎
95. @cyriux Same meaning, different wording "SEMI-FLEX" "EXCHANGEABLE" LEGACY API NEW API
96. @cyriux Number of seats available vs. List SEATS: 2 SEATS: [‘A12’, ‘A13’] LEGACY API NEW API
97. @cyriux Number of calls for the same use-case 3 CALLS 2 CALLS LEGACY API NEW API
98. 😮💨 Lots of discrepancies!
99. FULL AI APPROACH 😎
100. 100% AGENTIC AI LEGACY API NEW API MCP MCP MCP
101. The Verdict
102. 100% AI approach ✅ It works...: - Simple: 1 prompt does it all - No code
103. 100% AI approach ✅ It works... but: - Very slow 🐢 - Expensive 💰 - Not repeatable / reliable
104. ALTERNATIVE?
105. AI-GENERATED CODE 😎
106. AGENTIC AI + MCP CODE LEGACY API NEW API AI-POWERED GENERATION PHASE
107. AGENTIC AI + MCP CODE LEGACY API NEW API AI- GENERATED CODE ASSERTIONS LEGACY API NEW API
108. LEGACY API NEW API AI- GENERATED CODE ASSERTIONS AI-GENERATED CODE RUNS WITHOUT AI
109. ✅ Very fast! (Totally cheap) ✅ 100% Deterministic & 100% reliable once generated ⚠ Need to re-generate everytime the API changes Offline AI-approach
110. 👉 The Perfect Balance
111. 👉 Sometimes AI is better offline, generating code The Insight
112. Are AI projects like any other project?
113. SUCCESS = delivering what’s planed, on-time?
114. @cyriux @arollafr I was expecting a 100% AI solution…
115. @cyriux @arollafr I have to admit: the hybrid approach is the best one!
116. I tended to expect / envision specific solutions (from what I’ve learned)… As a manager:
117. But it didn't work that way And there were better alternatives! Surprise!
118. But it didn't work that way And there were better alternatives! Surprise! NOT A FAILURE!
119. But it didn't work that way And there were better alternatives! Surprise! IT WAS JUST BETTER!
120. Expect to be surprised! AI is NOT Intuitive
121. If we imagine too precisely: - What it could do - How it should work The Trap
122. If we imagine too precisely: - What it could do - How it should work The Trap WE'RE WRONG!
123. Too much certainty = constrain results & miss opportunities ULTRA IMPORTANT!!!
124. Accept to just LET GO
125. @cyriux Cultural impact Focus
126. With something new…
127. "Yeah but it won't work here!" "We're special!" "Our context is way too complicated!" The Common Objections
128. "Yeah but it won't work here!" "We're special!" "Our context is way too complicated!" The Common Objections FIRST STEP: MAKE SOMETHING THAT WORKS HERE!
129. "But it may fail!?"
130. "But it may fail!?" SURE: LET’S HAVE SOMEONE ELSE TAKE THE RISK!
131. Our experts are trapped in business as usual.
132. Our experts are trapped in business as usual. LET’S DELEGATE!
133. Then share results.
134. Then share results skills.
135. ✅ Technical Feasibility
136. ✅ Technical Feasibility ⬛ Cultural Feasibility
137. Knowledge Transfer
138. Knowledge Transfer Co-construction
139. "From now on, you do it. I’m just here to help."
140. ✅ Technical Feasibility ✅ Cultural Feasibility
141. WRAPPING UP
142. @cyriux AI PROJECT == COMPLEX PROJECT: EXPECT TO BE SURPRISED!
143. @cyriux AI: STILL COUNTER-INTUITIVE YOUR STARTING POINT DOESN'T PREDICT WHAT COMES NEXT
144. Also…
145. AI: very short trial cycles
146. Prompt are accessible: non-tech or low-tech profiles can contribute
147. @cyriux SHOCKING IDEA:
148. @cyriux SHOCKING IDEA: PROTECT SLACK TIME FOR DISCOVERY
149. @cyriux TO GET THE BEST OF AI… RECONSIDER HOW YOU APPROACH YOUR PROJECTS
150. EMBRACE UNCERTAINTY!
151. @cyriux https://www.linkedin.com/in/martraire Thank you! arolla.fr https://www.linkedin.com/in/thomas-nansot