Insights on enterprise AI, on-prem deployment, and regulatory compliance.

Written for the people who'll have to defend this in an audit.

₹1.78 cr ₹42.6 L CLOUD OWNAI 76% SAVINGS
Cost Analysis

The real 5-year TCO of ChatGPT Enterprise for 50 users in India.

A line-by-line breakdown: per-seat cost, INR conversion exposure, output-token inflation, and the per-call costs nobody quotes upfront.

APACHE 2.0 no flow-down no AUP §2.c LLAMA 3 CL "Built with Llama" flow-down AUP
Data PrivacyTechnical

Apache 2.0 vs Llama licence: what Indian enterprises need to know.

Why Llama 3's AUP §2.c, Mistral MRL's redistribution clause, and Gemma's remote-restriction clause matter the moment you ship a paid product.

M4 PRO 64 GB 14–18 tok/s ₹2.4 L · 1–5 users L40S 48GB · 1U ₹8 L · 15–50 users
Technical

On-prem LLM deployment: Mac Mini vs GPU server for 32B models.

When the M4 Pro is enough; when you actually need an L40S. Benchmarks on Qwen 3 32B at 8K context across three hardware tiers.

§ 11(g) escrow § 16(m–o) audit rights § 17(8) 6-hour clock RBI IT OUTSOURCING MD 10 APRIL 2023
RegulatoryNBFC

RBI IT Outsourcing Direction: what it means for AI vendors.

Section 11(g) source-code escrow, Section 16(m–o) audit rights, Section 17(8) 6-hour incident reporting — clause-by-clause for cloud LLM and on-prem AI vendors.

FINE-TUNE · LoRA Behaviour Tone Vocabulary RAG · Qdrant Citations Freshness Source bound use both
Technical

Fine-tuning vs RAG: when each makes sense for regulated data.

Behaviour shaping vs citation grounding. Why the answer is almost always "both" — and how to size them against your data churn rate and audit obligations.

Explain- ability Audit- ability Trace- ability RBI FREE-AI · AUG 2025
RegulatoryNBFC

Reading the RBI FREE-AI framework as a vendor.

What the August 2025 consultative paper signals about the binding guidance to come — and which vendor positions survive the hardening.

A L C O A+ C C E A ALCOA+ · SCHEDULE M (2024)
RegulatoryPharma

Schedule M revised 2024: what ALCOA+ means for AI outputs.

Attributable, Legible, Contemporaneous, Original, Accurate — and the four plus-attributes. What each clause demands when the "operator" is an LLM, not a person.

CUSTOMER LAN LLM India · on-prem US CLOUD DPDPA risk
Data PrivacyRegulatory

Data locality is a feature, not a tax.

The argument that data-residency requirements slow AI adoption assumes the cloud is the cheap path. For regulated workloads, on-prem is the cheap path — and the cost of cloud's optionality is the lawsuit you didn't see coming.

No posts in this category yet.

View all posts →