Qualitative Analysis of South Asian Cities

Researchers and UX teams studying cities like Mumbai, Dhaka, and Kolkata face a recurring problem: rich, relational urban practices don’t fit neat datasets. The Interactions ACM piece “Why AI Cannot Learn South Asian Cities” (posted August 19, 2025) documents how informally encoded systems, the ~5, 000 dabbawalas delivering >200, 000 meals daily, local festivals that reshuffle streets, and oral/local taxonomies, resist standard machine learning inputs. This post shows how to run a rigorous qualitative analysis of South Asian cities that respects local knowledge rather than forcing it into unsuitable models. You’ll learn a practical workflow (collection → transcription → custom coding → cross-segment synthesis), concrete measures to preserve vernacular categories, and how Evidano (www.evidano.com) speeds each step with secure transcription, multilingual normalization, thematic and cross-segment analyses, and visualizations that surface co-occurrence and hierarchical codes. Use these steps to turn episodic ethnographies and informal datasets into decision-ready insights without erasing the social logic the source article highlights.

Fast take: What the original piece shows

Nusrat Jahan Mim’s August 19, 2025 article argues that AI and standard data-hungry models misread South Asian cities because much of their coordination is informal, embodied, and cyclical. Read the original: www.interactions.acm.org/blog/view/why-ai-cannot-learn-south-asian-cities.

Concrete examples: a 19-year-old dabbawala (Junior Sharruk) handles 34 clients without GPS; ~5, 000 couriers deliver >200, 000 meals daily.
Key claim: these cities are not 'data-poor' but 'unreadable' to models trained on Western, codified datasets.
This creates a different research problem: how to analyze qualitative, vernacular, and relational data at scale without flattening it.

Findings snapshot

Date	Metric / Item	Value / Example	Source	Implication
Aug 19, 2025	Article published	Interactions ACM post by Nusrat Jahan Mim	www.interactions.acm.org/blog/view/why-ai-cannot-learn-south-asian-cities	Framing: cities shaped by informal knowledge resist ML legibility
Contemporary	Dabbawala network size	≈ 5, 000 couriers	Article	High performance from non-digital coordination
Contemporary	Daily deliveries	> 200, 000 tiffins/day	Article	Scale exists outside digital traces
Contemporary	Individual example	Junior Sharruk, age 19, 34 clients	Article	Embodied route knowledge; semiotic codes

What happened (plain English)

The article documents how South Asian cities run through overlapping temporalities and social infrastructures: analog logistics (dabbawalas), app-driven platforms (Swiggy, Zomato), religious and festival-driven spatial shifts (Jumma prayers in Dhaka, Durga Puja in Kolkata), and informal markets that are rarely captured in official records.

AI systems rely on stable, labeled, repeated patterns; the article shows many urban practices are episodic, relational, and locally coded.
When models are trained on Western-style, sensor-rich datasets they can misrepresent or erase vernacular categories and accountability mechanisms.
The result for analysts: missing variables, wrong inferences, and interventions that mismatch local governance and livelihood logics.

Implications for researchers and teams

For qualitative researchers & ethnographers

Treat verbal and embodied knowledge as primary data: record interviews, extract narratives, and preserve local taxonomies rather than forcing prebuilt labels.

Use cross-segment analysis to compare festival days vs. routine days, platform-covered corridors vs. analog corridors.

For UX and product teams

Don’t assume platform telemetry represents population behavior. Validate design decisions with field narratives and participatory mapping.

Prototype features that augment, not replace, local coordination (e.g., lightweight messaging that complements dabbawala codes).

For policy and urban planners

Design interventions that acknowledge cyclical, communal uses of space (religious gatherings, temporary pandals) and consult informal stakeholders.

Use qualitative synthesis to surface accountability networks that aren’t in formal registries.

Do more, faster with Evidano

Ingest mixed, vernacular inputs

Problem: data is in interviews, field notes, newspaper reports, and local scripts.

Evidano solution: import transcripts, reports, and scraped social posts; apply transcription and translation with custom dictionaries for Devanagari, Bengali, Urdu terms to preserve local categories.

Respect local categories while scaling analysis

Problem: pre-built taxonomies flatten vernacular codes.

Evidano solution: iterative codebook import + AI-assisted coding that proposes themes but preserves human-edited codes; hierarchical codes→subcodes keep nuance.

Compare moments and segments reliably

Problem: festival vs. routine days or app-users vs. informal-network users are hard to compare at scale.

Evidano solution: cross-segment and frequency analyses highlight which themes spike during Durga Puja, Jumma, or peak delivery hours.

Make insights explorable and secure

Problem: stakeholder buy-in needs clear evidence; data security matters for informal workers.

Evidano solution: clickable quotes, co-occurrence networks, word clouds, and encrypted storage; Evidano does not use customer data to train third-party models.

Practical 7-step workflow you can run this week

A quick run-book to turn field material into decision-ready insight.

1) Collect: Record interviews, field notes, and local social media posts; gather existing reports (e.g., ethnographies, vendor lists).
2) Transcribe & translate: Use Evidano transcription with a custom dictionary for local place and ritual names; enable PII redaction if needed.
3) Import: Upload transcripts and documents into Evidano and tag source metadata (date, neighborhood, event).
4) Seed codebook: Start with open coding on a representative subset; lock vernacular codes that must be preserved.
5) Run thematic + frequency analysis: Let Evidano surface common themes and their prevalence by segment (festival vs. non-festival; app user vs. informal courier).
6) Visualize & validate: Generate co-occurrence networks and share clickable quotes with local collaborators for validation.
7) Synthesize: Produce a short stakeholder brief with recommended, context-sensitive interventions.

Conclusion: From critique to method

The Interactions ACM piece reframes South Asian cities as systems that are often unreadable to standard AI. That’s an analytic opportunity: build workflows that surface relational knowledge instead of erasing it.

If you study urban informality, run the workflow above in Evidano to preserve vernacular categories, compare segments, and produce visual, stakeholder-ready outputs. Learn more or start a pilot at www.evidano.com.

Keep reading

Browse all articles