Feasibility Forecast

This document presents the feasibility analysis for migrating ERPNext from Python to Go, incorporating 3 weeks of observed intern velocity data from 7 teams (40-50+ interns). The core question: Can 3-5 interns migrate ERPNext in 3 months with unlimited AI credits?

The Scale of the Challenge

Dimension	Count	Migration Implication
Python LOC	316,679	Core business logic to translate
JavaScript LOC	73,932	Frontend to rebuild or replace
Unique doctypes	521	Each needs model + logic + tests
API endpoints	768	Each needs equivalent in target language
Reports	~177	Each needs query + formatting logic
Shared controllers	16 files, 23,212 lines	Deep inheritance; most complex to migrate
Modules	21	Each with cross-module dependencies
COCOMO estimate	~216 person-years	Traditional development effort

A full ERPNext migration requires Frappe framework replacement, 521 doctype translations, 768 API endpoints, 177 reports, 73K lines of frontend, regional compliance, data migration tooling, and third-party integration points.

Observed Velocity Data

Reference Implementation (Experienced Engineer)

Iteration	Python Source	Go Output	Tests	Calendar Time
Mode of Payment	66 lines	518 lines	19	~1 day
Tax Calculator	1,334 lines	1,240 lines	24	~1 day
GL Entry Engine	878 lines	2,608 lines	32	~1 day

Rate: ~760 lines of Python source per day per person. Code expansion factor: ~1.9x.

Intern Team Velocity (3 Weeks Observed)

Team	Active Contributors	Primary Output	Velocity Assessment
Epsilon	3	CI-ready parity tests, Groq integration, 6 PRs	High
Zeta	1 (solo)	100% tax parity, RAG, Streamlit, TS SDK	High
Beta	2	VS Code extension (1,700 LOC), Redis cache, 7 PRs	High
Gamma	2	GraphRAG eval, Sales Invoice 89.1% coverage	Medium-High
Alpha	4	RAG + Streamlit + submodule execution	Medium
Delta	3	Python-Go transpiler, 3-phase analyzer	Medium
Eta	2	RAG evaluation pipeline only	Low

Velocity Discount Factors

Factor	Original Estimate	Observed Reality
Intern vs experienced engineer	0.3x	0.2-0.5x (wide variance by team)
Learning curve (first 2-4 weeks)	0.5x during ramp-up	Confirmed — Weeks 1-2 were onboarding
AI tooling acceleration	3-5x	3-5x confirmed
Review and rework cycles	0.7x	0.5x (no systematic code review for most teams)
Coordination overhead	Not estimated	0.6x (significant for 40+ interns)
API quota / tooling friction	Not estimated	0.8x (recurring blocker)
Net effective velocity	~0.5-1.0x	~0.3-0.8x of experienced engineer

Scenario Analysis

Scenario A: Full ERPNext Replacement (3 months, 3-5 interns)

Verdict: Not feasible.

Work Item	Effort Estimate
521 doctypes	8-12 months
Framework replacement	6-12 months
768 API endpoints	2-3 months
177 reports	1-2 months
Frontend (73K JS lines)	4-6 months
Regional compliance	2-3 months
Data migration	1-2 months
Integration testing	2-3 months
Total	26-43 months of work

Even with unlimited AI credits and 5x acceleration, architectural decisions, domain understanding, and integration testing cannot be compressed below 5-6 months for a skeleton with major gaps.

Scenario B: Core Accounting Engine (3 months, 3-5 interns)

Verdict: Achievable — supported by observed intern data.

Module	Doctypes	Weeks
GL Engine (done)	1	0
Tax Calculator (done)	1	0
Mode of Payment (done)	1	0
Account Master	~10	2
Journal Entry	~5	2
Payment Entry	~8	3
Sales Invoice	~15	3
Purchase Invoice	~12	2
Bank Reconciliation	~5	1
Cost Center / Dimensions	~8	1
Budget	~5	1
Fiscal Year / Period	~5	1
Total	~75	~16 weeks

Evidence from 3 weeks of intern work:

Zeta independently replicated tax calculation parity (approach is teachable)
Epsilon built CI-ready parity tests (testing methodology validated)
Gamma’s golden dataset reveals accuracy gaps in cross-module queries
4 of 7 teams built working Streamlit UIs for verification

Go vs Kotlin Comparison

Dimension	Go	Kotlin
Type system	Simple, explicit	Rich (nullable, sealed classes)
Error handling	Explicit returns	Exceptions + Result type
Concurrency	Goroutines + channels	Coroutines
ORM ecosystem	sqlx, GORM, Ent	Exposed, Spring Data
Web framework	Chi, Echo, Fiber	Ktor, Spring Boot
AI training data	Moderate	High (Java/Kotlin ecosystem)
ERP domain fit	Good for microservices	Better for full framework replacement
Existing work	3 iterations complete	None

Recommendation: Go for microservice-style extraction (proven patterns). Kotlin if building a full Frappe framework replacement (Spring Boot ecosystem provides ORM, permissions, workflow out of the box).

Cost Projection (3-Month Phase 1)

AI Credits

Usage	Monthly Estimate
Claude API (code translation, review, test gen)	$2,000-5,000
GPT-4 API (alternative/comparison)	$1,000-3,000
Embedding/indexing	$200-500
MLflow + evaluation	$100-200
Total AI per month	$3,300-8,700

Human Cost

Role	Count	3-Month Total (INR)
Interns	5	3.75-6.0 lakh
Senior architect (part-time)	1	1.5-3.0 lakh
Total		5.25-9.0 lakh (~$6,300-10,800 USD)

Total Budget

Category	Low	High
AI credits	$10,000	$26,000
Human cost	$6,300	$10,800
Infrastructure	$500	$1,500
Total	$16,800	$38,300

Risk Assessment (Updated with Observed Data)

Risk	Probability	Impact	Evidence
Framework replacement underestimated	High	Critical	No team attempted framework-level work
Controller inheritance complexity	High	High	Gamma’s golden dataset: 1/5 on GL precision
Self-reported accuracy inflation	High	High	Alpha ~70% vs Epsilon/Zeta “100%” claims
Quality variance across teams	High	High	5x quality gap between best and worst teams
Intern attrition / skill gaps	High	High	Eta inactive 2+ weeks, exam conflicts
API quota exhaustion	Medium	Medium	Multiple teams reported rate limits
Regional compliance gaps	Medium	High	Only Indian GST attempted
Data migration breaks	Medium	Medium	Not yet attempted

What AI Actually Compresses

Phase	Compressible?	AI Impact
Understanding legacy code	Yes (5-10x)	High
Mechanical translation	Yes (10-50x)	High
Test generation	Yes (3-5x)	Medium-High
Architecture decisions	No	None — needs human judgment
Domain validation	Partially	AI drafts, human verifies
Framework design	No	None — needs senior engineering

The Honest Assessment

For a full ERPNext equivalent:

Traditional (no AI): ~5-8 years with a team of 10
AI-assisted: ~12-18 months with a team of 8-12
AI-assisted (interns only, core modules): ~3-6 months with a team of 5

The right question: “What is the minimum viable extraction that delivers production value?” For most businesses using ERPNext, the accounting engine IS the product. A Go accounting engine with clean APIs, extracted with full parity testing, deployed alongside ERPNext via Strangler Fig — that is achievable in 3 months with a focused team.

Key recommendation from observed data: Optimize for team quality over team size. A focused pod of 3-5 high-engagement interns with enforced PR review, provisioned API keys, and weekly parity checkpoints outperforms a 40-person cohort with variable commitment.