Why 95% of AI Projects Fail, and What "Solve First, Automate Later" Fixes
95% of AI projects fail to reach production. Not because the models are not smart enough. They fail because organisations automate broken workflows and call it transformation.
scroll
Oslo, Norway
Rishi Raj
Manglesh
Solve First. Automate Later.
Building the AI layer for accounting and insurance. Twenty years across India, Asia, and Norway. That depth is the product.
About
Intelligence built for the systems accounting and insurance already run.
Since 2021 I have been building Hundred Solutions, and in 2026 DVERSI, an AI platform that connects to the systems accounting and insurance companies already run, without replacing them. Not a chat interface. An intelligence layer via MCP that builds on what exists and compounds over time.
The twenty years before that are the product, not the background. Running life & pension insurance and banking systems across India, Asia, and Norway gave me domain depth that generic AI vendors do not have.
2021 – PresentFounder, Hundred Solutions · Oslo
2026 – PresentFounder, DVERSI
Published 2026
BookAI Orchestration: The Definitive Guide
ResearchDarshana: Six-School Framework for LLM Orchestration
Background20 years · Insurance & banking · India, Asia, Norway
Published Work
AI Orchestration:
The Definitive Guide
How to Build Automation That Actually Works
Rishi Raj Manglesh · Published by Hundred Solutions AS · May 2026
Most organisations bolt AI onto legacy tools without understanding why the process exists. This book is a diagnostic before an implementation. Every framework earned through production, every warning earned through failure. Accounting and insurance operators will recognise the problems. The patterns apply across industries.
click cover to flip
Research
Darshana: six schools of Indian epistemology as LLM engineering layers
Two open-source repos · ~5,000 generations · validated against equal-sophistication generic controls · cross-judged by GPT-4o. The core finding: the same technique performs 0% or 82% depending on where in the stack it is applied.
Preprint · 2026
Darshana: A Six-School Framework for Large Language Model Orchestration and Training
The six classical Indian philosophical schools (Nyaya, Vaisheshika, Samkhya, Yoga, Mimamsa, Vedanta) each map to a specific layer of the LLM engineering stack. Layer assignment is the critical variable: Mimamsa scored 0% as a runtime system prompt but 82% applied as a query rewriter. Same technique, different layer.
Results vs generic controls
83%
Vritti: Epistemic self-classification
60% cross-judge (GPT-4o)
93%
Nyaya: Tool routing via 4 Pramanas
70% search reduction
82%
Mimamsa: Query rewriting (6 Lingas)
67% cross-judge
71%
Vaisheshika: Knowledge ontology (7 Padarthas)
vs generic entity ontology
82%
Vedanta: Output synthesis
63% cross-judge
Work
What I’m building
DVERSI
AI Platform
The AI layer for accounting and insurance systems. DVERSI connects via the Model Context Protocol. The customer keeps their existing stack entirely, and adds coworkers, workflows, and compliance enforcement on top. Eight to nine weeks to a working pilot. Five to thirty times cheaper than a replacement platform.
- Customer's existing systems preserved
- Domain-native: accounting & insurance
- Entity-centric knowledge graph
Hundred Solutions
Advisory & Consulting
The activation and services partner for organisations adopting DVERSI. We deliver the scoping, integration, custom UI, and compliance work that turns the platform into a working system for a specific team, in eight to nine weeks, not eighteen months.
- Activation engagements & vertical extensions
- Strategic advisory retainer
- Oslo, Norway
Writing
Notes from the field
The AI Industry Has an Operating System, and Most Builders Don't See It
Every AI startup in 2026 is building an application. Almost none are asking the question that actually determines survival: what operating system are they running on, and what happens when it changes?
What 2,000-Year-Old Indian Epistemology Taught Us About Building Better AI
We took the six classical Indian schools of philosophy and asked whether 2,000-year-old frameworks for organising knowledge actually improve modern AI systems. Across 4,400 experimental generations, the answer was unambiguous.
Talks & Conversations
The conversation before the decision
Most of my speaking happens in rooms where a decision is being made: a board evaluating an AI investment, a leadership team committing to a direction, a technical group designing something they will live with for years.
I bring twenty years inside accounting and insurance systems and published research on LLM engineering. No product pitch. If an independent perspective is useful before you commit, reach out. Most conversations start with one question.
Start a conversationAI Advisory
Where AI works in financial operations and insurance, and where it silently fails.
Fractional CTO
Senior architecture and technology leadership without a full-time hire.
Solve First, Automate Later ™
Proprietary methodology from the book: what to automate, in what order, and why.
Ancient Epistemology & Modern AI
Classical Indian knowledge frameworks mapped onto LLM engineering layers.