SelfMate : An AI shopping assistant for Walmart Ecosystem
The AI Shopping Assistant That Answers Before You Finish Typing
Meet an AI shopping assistant that sees, thinks, and acts in real time—finding the right product, checking local stock, applying the best price, and explaining why, often in under a heartbeat.
Hero stats:
<200 ms cached replies
10k+ concurrent users
95%+ recommendation accuracy (stress‑tested in production)
Engineered for scale: microservices, Gemini 2.5 Pro multi‑model, and a polyglot data layer tuned for retail .That can serve on all platform : shopping platform, kisosk and as standalone chatbot app .
User Pain Hold and Why this Problem Statement :
The user pain is clear: keyword boxes and endless catalogs slow decisions, especially in‑aisle where price, diet, and local stock all collide; expectations set by Amazon’s constant tech leaps demand effortless, guided discovery, so Walmart must turn its unmatched offline footprint into a digital advantage.
This problem statement targets that gap: a lightweight assistant that elevates the in‑store experience for shoppers and a companion extension for associates speeding availability checks, suggesting substitutes, and accelerating pick/pack reducing abandonment while converting Walmart’s physical scale into a durable edge.
System at a Glance :
Figure 1. System Architecture Tool Used→
Figure 2. Data Flow Diagram
How it works (in 5 beats) :
Intent lock: Domain prompts route queries to “search, compare, refine” with schema‑locked tools.
Dual‑tier routing: Tier 1 handles the essentials instantly; Tier 2 adds availability, alternatives, and trends when context demands.
Search + semantics: Typesense narrows fast; Qdrant re‑ranks by meaning and visuals; fused for precision.
Reality check: Live inventory and regional pricing blend into the final rank so answers match shelf truth.
Personal touch: Profile, behavior, and store affinity shape a personalization score alongside relevance.
User Experience:
Figure 3. User Flow Diagram
Start with scan, snap, ask: users scan a shelf QR/tag, snap an item photo, or speak/type “gluten‑free cereal under INR 100 near 75001,” the assistant locks intent and fires Tier 1 search to filter by price and diet, fuses semantic matches via Typesense and Qdrant, escalates to Tier 2 for live availability, pickup/delivery and nearest‑store distance, then returns a tight “Your Smart Picks” panel with the best option (₹85), why‑this choice, and healthier alternatives—one smooth, real‑time glide from question to add‑to‑cart .
Lesson Learned :
Accuracy and Relevance: We fixed hallucinations and relevance issues by using a hybrid search approach (Typesense + Qdrant re-ranking) and ensuring that all responses are grounded in real data, with built-in retries for failures.
Latency and Efficiency: We dramatically reduced latency spikes by implementing parallel processing, short-term caching, and setting strict time limits for each step. We also managed token costs by using a lower model temperature and compressing memory.
