Global AI & digital engineering partner. Agentic AI systems, RAG apps, unified BOT platforms, cloud & DevOps, mobile/web, Web3 & IoT. Enterprise to startup. DM.
India •SG • UK • US • UAE • SA
Based in India
- Tweets 28,106
- Following 2,898
- Followers 2,407
- Likes 14,592
Jan 1, 2027: AI systems will face tougher disclosure expectations.
Can your AI show its receipts?
mobiloitte.com/contact-us?ut…
#AIGovernance #AIRegulation
17
AI is easy to demo. Production is where architecture gets real.
Data ownership. System boundaries. Approvals. Rollback. Observability.
Treat the handoffs as part of the product—not plumbing.
mobiloitte.com/contact-us?ut…
#EnterpriseAI #SoftwareArchitecture
1
27
2009: a mobile-app team in Okhla explaining what an “app” was. 2026: six offices, clients in 70+ countries, 5,000+ projects (figures: mobiloitte.com). The constant: the gap between demo and deployed — and the discipline that closes it.
mobiloitte.com/about-us/
18
Security shouldn’t be a final QA checkbox.
Ask one question: Which phase is hardening?
If it has no clear phase or exit criteria, security may still be an add-on.
mobiloitte.com/contact-us?ut…
#CybersecurityAwarenessMonth
12
AI in banking doesn’t have to make the decision to add value.
It can read the documents, structure the data, flag exceptions—and leave judgement with people.
More: mobiloitte.com/industries/
#BFSI
9
19.8M people died from cardiovascular disease in 2022 (WHO). Over 80% of these deaths are linked to risk factors that can be prevented or managed.
AI is helping doctors spot heart disease sooner.
#WorldHeartDay #DontMissABeat ❤️
17
80%+ of world trade by volume moves by sea, carried by 2.57M seafarers.
In 2026, IMO adopted the first global code for autonomous ships. The AI era at sea is starting with rules and with people.
Happy #WorldMaritimeDay
210
In regulated workflows, the model is not the deliverable.
The auditable process is.
Outputs that cannot be explained, versioned, or reproduced are unusable in banking, insurance, and healthcare.
mobiloitte.com/contact-us?ut…
38
Fine-tuning is not for missing facts.
Its real signal is behavior variance you cannot eliminate.
If the model behaves correctly most of the time, but not reliably enough for production, tuning may be a candidate.
mobiloitte.com/contact-us?ut…
31
Prompting first.
Then retrieval.
Then fine-tuning.
That is the escalation path.
Most teams underinvest in prompt design, conclude the model cannot do it, and jump two rungs too early.
mobiloitte.com/contact-us?ut…
28
Before fine-tuning, run the diagnosis:
Facts wrong = retrieval problem.
Behavior inconsistent = fine-tuning candidate.
Instructions ignored = prompting problem.
The right fix starts with the right diagnosis.
mobiloitte.com/contact-us?ut…
22
One AI project is an achievement.
Repeated AI delivery is a capability.
That requires ownership, reusable foundations, honest evaluation, governance, monitoring, and a real operating model.
mobiloitte.com/contact-us?ut…
23
AI adoption is not automatic.
A technically successful system with no users delivers nothing.
Production AI needs workflow change, business ownership, training, rollout planning, and trust.
mobiloitte.com/contact-us?ut…
18
Most enterprises do not need to fine-tune an LLM.
They reach for it too early and spend months arriving where retrieval would have started.
Fine-tuning is powerful — but only for a narrow band of problems.
mobiloitte.com/contact-us?ut…
12
AI use-case selection is the cheapest risk control in the whole programme.
Pick use cases with a measurable problem, workable data, manageable integration, realistic accuracy bar, and real adoption path.
mobiloitte.com/contact-us?ut…
26
The best AI production playbook has five phases:
Problem definition.
Data assessment.
Honest POC.
Production build.
Deployment, adoption, and operation.
Skip one and risk the whole rollout.
mobiloitte.com/contact-us?ut…
24
The gap between AI pilot and AI production is not one step called deployment.
It is seven streams:
Data.
Reliability.
Integration.
Scale.
Governance.
Operations.
Adoption.
mobiloitte.com/contact-us?ut…
25
The POC trap is simple:
The proof of concept works, so people assume the product is nearly done.
But the POC only proved feasibility.
Integration, governance, reliability, adoption, and scale still remain.
mobiloitte.com/contact-us?ut…
27
The real test of AI ethics is not the policy document.
It is how the system treats users with the least power to detect, contest, avoid, or recover from harm.
mobiloitte.com/contact-us?ut…
27
Designing for vulnerable users is not a trade-off.
Clear communication, accessibility, robustness, and recourse improve the system for everyone.
mobiloitte.com/contact-us?ut…
46
A clean AI pilot can look impressive.
Production is different.
Production has messy data, edge cases, real users, security, integration, governance, and scale.
That is where the real work begins.
mobiloitte.com/contact-us?ut…
22
AI pilots often die before production for one simple reason:
They were launched to “use AI,” not to solve a measurable business problem.
No number to move means no bar for success.
mobiloitte.com/contact-us?ut…
28
Most enterprise AI pilots do not fail because the model is weak.
They fail because the business problem, data, integration, adoption, governance, or production path was never properly planned.
mobiloitte.com/contact-us?ut…
35
Consent does not fix every AI ethics problem.
If a user cannot realistically understand the interface, the consent is not meaningful.
Design must carry the duty.
mobiloitte.com/contact-us?ut…
48
AI systems often fail vulnerable users in four ways:
Typical-user design.
Thin training data.
Limited testing.
Weak recourse.
These failures compound.
mobiloitte.com/contact-us?ut…
33
The same person can be vulnerable to one AI system and not another.
Nothing about the person changed.
The system relationship changed.
mobiloitte.com/contact-us?ut…
48
A person becomes algorithmically vulnerable when AI failure can cause serious harm and they have limited ability to detect, contest, avoid, or recover from it.
mobiloitte.com/contact-us?ut…
48
“Vulnerable populations” is used often in AI ethics.
But vulnerability is not a fixed label on people.
It is created by the relationship between a person and a system.
mobiloitte.com/contact-us?ut…
76
A duty of care is not a disclaimer.
Not a consent screen.
Not a values statement.
It must show up in design, evaluation, deployment, governance, and recourse.
mobiloitte.com/contact-us?ut…
38
AI systems often fail vulnerable users in four ways:
Typical-user design.
Thin training data.
Limited testing.
Weak recourse.
These failures compound.
ow.ly/Gm3850ZeaJV
5
Situational vulnerability can affect anyone.
Illness. Financial shock. Bereavement. Displacement. Crisis.
AI systems must work for users on their worst days.
mobiloitte.com/contact-us?ut…
13
If vulnerability is treated as a deficiency in the person, responsibility disappears.
If it is created by the system relationship, builders have a duty to reduce it.
ow.ly/sE4q50ZeaGw
9
The same person can be vulnerable to one AI system and not another.
Nothing about the person changed.
The system relationship changed.
ow.ly/qI7J50Zea8j
6
A person is algorithmically vulnerable when AI failure can cause serious harm and they have limited ability to detect, contest, avoid, or recover from it.
ow.ly/RosJ50Zea2o
7
“Vulnerable populations” is used often in AI ethics.
But vulnerability is not a fixed label on people.
It is created by the relationship between a person and a system.
ow.ly/4HC050Zea1p
8
Goal-directed
Plans
Uses tools
Has bounded autonomy
Fastest filter: when it finishes, did anything actually change in a live system?
mobiloitte.com/contact-us?ut…
#Mobiloitte #AgenticAI #AIAgents #EnterpriseAI #GenAI
7
Fine-tuning is not for teaching facts.
Facts change. Facts need citations. Facts need updates.
That is what retrieval is for.
Fine-tuning is for behavior.
ow.ly/xiUv50Z9VEU
6
A prototype gets outputs wrong.
The usual reaction: “Fine-tune it on our data.”
Often, that is the wrong move.
First ask: is the issue knowledge, behavior, tools, or evaluation?
ow.ly/K3N850Z9VuN
10
Most enterprises that think they need to fine-tune an LLM probably do not.
Fine-tuning is powerful, but it is not the first lever.
Prompting, RAG, and tools should be tested first.
ow.ly/Wx4j50Z9VfO
6
Co-lending is operationally complex.
Two regulated entities. One borrower. Shared books. Different policies.
AI can reduce friction across eligibility, underwriting, disbursement, servicing, and reconciliation.
Visit: ow.ly/ISGt50Z8FJE
7
Digital lending in India is one of the clearest AI opportunities.
Underwriting, fraud checks, document AI, vernacular communication, and servicing can improve when built within the regulatory frame.
Visit: ow.ly/Cxnc50Z8FHc
12
AI for NBFCs and banks is not just automation.
It must work inside RBI expectations, customer protection, model risk, digital lending rules, and auditability.
Visit: ow.ly/u5Eb50Z8FCk
9
Indian financial services has built powerful credit rails: Aadhaar, UPI, Digital KYC, Account Aggregator, ULI, DPDP, and RBI frameworks.
AI can unlock the next layer of value — if built with discipline.
Visit: ow.ly/OPTB50Z8ENs
15
Collections AI can improve outcomes only when built with discipline.
Right timing. Right channel. Vernacular support. Conduct guardrails. Escalation. Audit trails.
That is where AI becomes operationally useful.
ow.ly/2EPJ50Z8nCw
4
Co-lending is operationally complex.
Two regulated entities. One borrower. Shared books. Different policies.
AI can reduce friction across eligibility, underwriting, disbursement, servicing, and reconciliation.
ow.ly/FR2F50Z8nBM
3
Digital lending in India is one of the clearest AI opportunities.
Underwriting, fraud checks, document AI, vernacular communication, and servicing can improve when built within the regulatory frame.
ow.ly/4Mvx50Z8nAn
10
AI for Indian NBFCs and banks is not just about automation.
It must work inside RBI expectations, customer protection, model risk, digital lending rules, and auditability.
ow.ly/liwa50Z8nzg
4
Indian financial services has built a powerful credit infrastructure.
Aadhaar, UPI, Digital KYC, Account Aggregator, ULI, DPDP, and RBI frameworks have changed the game.
Now AI can unlock the next layer of value.
ow.ly/vJTS50Z8nwB
7
Enterprises do not need to rebuild everything for AI.
They need to sequence the foundation around the use cases that matter most.
ow.ly/qII550Z7z4f
4
AI-ready data platforms are not tools.
They are engineered layers for context, quality, access, lineage, retrieval, and feedback.
ow.ly/rJIk50Z7z3t
4