Building a Comprehensive LLM Platform for Food Delivery Services

swiggy · E-commerce · 2024 · source Agentic: no · promotional bias: med · summary-only (no fetched source) · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Food-delivery generative AI platform

Swiggy needed to explore generative AI across customer experience and operations while balancing hallucination risk, latency, data governance, ROI uncertainty, and the different requirements of image generation, review summarization, search, and support bots.

Constraints:

• source content could not be fetched, so this extraction is summary-only
• search requires around 100ms latency while chatbots can tolerate around 3 seconds
• food image generation must handle Indian dishes underrepresented in off-the-shelf models
• customer-facing conversational interfaces showed limited traction
• many initiatives required 3-4 months of iteration before viable ROI

Architecture

The summary describes a portfolio-style GenAI platform: Stable Diffusion with LoRA fine-tuning for catalog imagery, GPT-4 review summarization, RAG WhatsApp bots for restaurant partner support, neural search, governance and latency controls, and a prioritization framework that favors high-demand lower-risk use cases before higher-risk automation.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  genai-task-force[GenAI task force] --> demand-risk-framework[Demand risk framework]
  demand-risk-framework --> image-generation-pipeline[Image generation pipeline]
  demand-risk-framework --> review-summarizer[Review summarizer]
  demand-risk-framework --> rag-support-bot[RAG support bot]
  demand-risk-framework --> neural-search[Neural search]
  rag-support-bot --> governance-layer[Governance layer]
  neural-search --> governance-layer
  review-summarizer --> a-b-evaluation[A/B evaluation]
  governance-layer --> a-b-evaluation

Components

Data flow

Different Swiggy use cases flow through different model paths: catalog image tasks use Stable Diffusion and LoRA, review summarization uses GPT-4 prompts and A/B testing, restaurant support uses RAG over SOP content through WhatsApp, and search uses neural models optimized for real-time latency. Governance, hallucination mitigation, and data-security controls sit across the portfolio.

Why this architecture

Swiggy treated GenAI as a portfolio rather than one chatbot bet. Different use cases had different latency, risk, and ROI profiles, so the architecture mixed foundation models, fine-tuning, RAG, neural search, governance, and A/B testing instead of forcing one model path onto every problem.

Models used: Stable Diffusion · LoRA · GPT-4 · LLMs · neural search models

Scale

• The initiative began in early 2023.
  “Swiggy, India's leading food delivery platform, embarked on a comprehensive generative AI initiative in 2023 to enhance customer experience and operational efficiency.”summary, stated
• Review summarization A/B testing involved over 2,000 restaurants.
  “In A/B testing involving over 2,000 restaurants, they observed improvements in funnel metrics and reductions in cancellations and claims, attributed to enhanced expectation management.”summary, stated
• Many initiatives required 3-4 months of iteration before viable ROI.
  “Results included improved funnel metrics and reduced cancellations in A/B testing for review summarization, though the team acknowledged that many initiatives required 3-4 months of iteration before achieving viable ROI, and customer-facing conversational interfaces showed limited traction.”summary, stated

Alternatives considered (roads not taken)

Customer-facing conversational interfaces as a primary bet

Rejected because: The summary says food-ordering/table-reservation chatbots showed limited traction.

Off-the-shelf text-to-image models for all food categories

Rejected because: They struggled with Indian dishes, prompting LoRA fine-tuning.

Patterns

fine-tuning-adaptation (model-ops)

LoRA fine-tuning adapts image generation for Indian dish categories.

rag-retrieval (retrieval)

Restaurant partner support uses a RAG pipeline over SOP documents through WhatsApp.

eval-driven-development (evaluation)

Review summarization was evaluated through internal metrics and A/B tests.

latency-optimization (production)

Swiggy distinguishes search latency needs from chatbot latency tolerance.

Tradeoffs

Use-case prioritization

Chose high-demand, low-risk tasks first over higher-risk resolution-focused chatbots and process automation first.

Rationale: Early wins needed business value without excessive hallucination or automation risk.

Model strategy by latency

Chose customized LLMs for real-time use cases over GPT APIs for all use cases.

Rationale: Non-real-time tasks could use GPT, but search-like tasks required strict latency.

Gotchas (what bit them)

Hackathon demos did not reliably translate into ROI.

Root cause: Generative AI applications needed months of iteration and stakeholder expectation management.

Resolution: Use a demand-risk framework and patience around ROI validation.

Direct GPT usage created governance difficulties.

Root cause: External provider data use, PII, and security needed centralized control.

Resolution: Move toward third-party API providers and an internal platform layer with governance.

Evaluation

Swiggy used demand-risk prioritization, internal quality metrics, A/B tests, funnel metrics, cancellation/claim tracking, latency targets, and ROI iteration windows.

Metrics: 100ms search latency · 3 second chatbot tolerance · A/B funnel metrics · cancellations and claims · 3-4 month ROI iteration

Outcomes

• Review summarization improved funnel metrics and reduced cancellations in A/B tests.
• Catalog image generation needed LoRA adaptation for Indian dishes.
• Conversational interfaces showed limited customer traction.

The extraction is summary-only; the summary describes Swiggy as one of India's largest food delivery and quick commerce platforms.