Practical AI Nepal  ·  Frameworks

Four frameworks. One reading list for institutions.

The Problem-to-Pathway diagnostic, the Practical AI Stack, the Model Improvement Ladder, and the Inference Ladder. Each is short, defensible, and built to be used inside real workflows — not to sit on a slide.

§ Framework 01

The Problem-to-Pathway diagnostic.

Most AI projects fail not at deployment but at framing. This diagnostic forces the institution to name the problem before the tooling — and routes them to the appropriate response, which is often not a model.

A

Learning gap

The team doesn't yet understand what AI can and cannot do. → Briefing, workshop, AI literacy program.

B

Governance gap

The institution lacks the rules, review, and disclosure to deploy AI safely. → Policy work, governance toolkit, intake process.

C

Workflow gap

The current process is unclear or broken; AI would only automate confusion. → Workflow redesign first, AI second.

D

Evaluation gap

The institution doesn't know whether existing models work for its task or language. → Model evaluation, Nepali-context benchmark.

E

Adaptation gap

Models work for general tasks but miss domain or local nuance. → RAG, fine-tuning, or continued pretraining — in that order.

F

Inference gap

The model is fine; running it on a public API would breach data, cost, or sovereignty constraints. → Local inference, on-prem, or edge.

G

Implementation gap

Diagnosis is done; the institution needs a partner to actually build, deploy, and operate the system. → Implementation partner.

§ Framework 02

The Practical AI Stack.

Nine layers, top to bottom. Each is a separate question — and a separate point of failure. Practical AI is the cumulative answer.

01
Policyनीति

Rules, rights, standards, and responsibility. The legal and procedural baseline against which every AI deployment is judged.

02
Dataडाटा

Local datasets, consent, privacy, and quality. What the country has, what it can collect, and what it must protect.

03
Computeगणना

GPU and TPU access, cloud, and local infrastructure. The physical capacity the country can actually call on.

04
Modelsमोडेल

Open models, closed APIs, and domain adaptation. Which models the country chooses to evaluate, deploy, and adapt.

05
Inferenceअनुमान

Where AI runs and who controls it. The single most important governance decision once a model is in production.

06
Edgeकिनारा

Devices, schools, rural nodes, volunteer compute. Distributed participation that does not require central data flow.

07
Evaluationमूल्यांकन

Where models fail in the Nepali context — language, domain, code-mixing, OCR, public-service tasks.

08
Human Controlमानव नियन्त्रण

Review, approval, appeal, accountability. The humans who can stop, redirect, or override the system.

09
Implementationकार्यान्वयन

Real workflows, training, adoption. The point at which the stack becomes practice instead of intent.

§ Framework 03

The Model Improvement Ladder.

Use this every time someone asks, "Should we fine-tune?" Most of the time the answer is on a lower rung — and the lower rungs are cheaper, faster, and safer.

01

Prompting fix

Better instruction. Re-state the task, the role, and the constraints. The cheapest possible improvement.

02

Workflow fix

Better UI, review, and guardrails. Often the model is fine and the surrounding workflow is the actual problem.

03

RAG / grounding fix

Trusted local knowledge. Plug the model into the institution's own documents, policies, and Nepali-context data.

04

Fine-tuning fix

Consistent task behaviour. Use when the failure is reproducible across prompts and the same correction would always apply.

05

Continued pretraining fix

More Nepali and domain exposure. A serious investment, justified by serious gaps.

06

Foundational pretraining fix

Long-term national capability. The top of the ladder. Not the first answer.

Not every model limitation needs a new model. Some need better workflow, some need grounding, some need fine-tuning, and only some need pretraining.
— Practical AI Nepal · signature line
§ Framework 04

The Inference Ladder.

Where AI runs is a governance decision. Each rung trades convenience for control. The right rung depends on data sensitivity, accountability, cost, and continuity — not on what is fashionable.

01

Public API

Closed model, third-party server. Fastest to deploy, weakest control over data, vendor, and continuity.

02

Enterprise / private API

Same vendor, contracted controls. Better residency and audit terms; still a vendor dependency.

03

Hosted open model

Open weights on a managed cloud. Portable across vendors; still requires cloud trust.

04

Local GPU server

Open model on the institution's own hardware. Data stays local; institution carries operations.

05

On-prem / air-gapped

No outbound network. The choice for sensitive public-sector and regulated workloads.

06

On-device / edge

Phone, laptop, school node. Distributed participation; small models, minimal data flow.

§ Lines

Ten lines we repeat in public.

The phrasings we use in briefings, op-eds, and ministry meetings. Plain enough for a newspaper, defensible inside a procurement review.

L · 01

Nepal should not become only an AI consumer market.

L · 02

AI governance should not live only in PDFs.

L · 03

Practical AI means AI that institutions can actually use and govern.

L · 04

Closed APIs are useful; open models are strategic.

L · 05

Where AI runs is a governance decision.

L · 06

Not every model problem needs fine-tuning.

L · 07

Nepal needs compute strategy, not just AI tool adoption.

L · 08

Edge AI can let communities participate in Nepal's AI future.

L · 09

AI sovereignty is not isolation; it is optionality.

L · 10

Governance becomes real only when it changes workflows.