Agrim AI
A Shopify-integrated generative-AI e-commerce platform — a 10-microservice AI backend behind a WAF + Application Gateway on a segmented Azure VNet. Within it, I independently designed, built, and own two production microservices end-to-end:
Image Service Virtual Try-On · Floor · Wall Paint
A unified FastAPI service that consolidates three visual product-visualization capabilities into one containerized deployment — turning a person + garment, a room + tile, or a wall + paint into photorealistic previews before purchase.
- Consolidated three capabilities — Virtual Try-On, Floor Visualizer, and Wall Paint — into a single deployment, each isolated in its own module with graceful degradation: if one fails to initialize, the others keep serving, and
/healthreports per-capability readiness. - Generated three photorealistic views (front / side / back) per try-on via Gemini, with an "Identity Guard" prompt that locks the customer's face, hair, and body and treats the garment purely as source — eliminating the common identity-swap failure.
- Built a five-layer safety gate (person-presence, minor detection, nudity, garment validity, lighting) with fail-closed safety checks and fail-open quality checks, plus a post-generation pass on the model's own output.
- Added confidence-based non-wearable filtering so décor/objects can't hallucinate into garments, with batched VLM prechecks (8 concurrent, verdict cached per product) so storefront grids never hammer the model.
- Engineered two-tier wall masking — MobileSAM point-prompted precision gated by a Segformer wall-class mask, with auto wall-detection over a probe grid and IOU dedup; floors use Segformer + a perspective-aware warp to re-tile a room.
- Cut try-on first-view latency ~29s → ~11s via progressive delivery, and used downscale-then-composite to keep Gemini floor renders around ~15-20s.
Integration Service API Gateway · Webhook Router
A configurable API-to-API gateway that decouples the platform from third-party services — clients register an integration once as a database "recipe", with no code changes or redeploys to onboard a new external API.
- Designed a recipe-driven gateway where each integration (endpoint, method, request template, expected response keys) lives as a MySQL row — zero-code onboarding of new third-party APIs per client.
- Wrote a
{{variable}}template engine supporting nested dotted paths and deep recursive leaf lookup, so one recipe adapts to varied upstream payload shapes without duplication. - Verified responses by deep-searching required keys and confirming each is actually filled (not null/empty), with automatic configurable retries and full failure-state capture (status, missing keys, upstream body) for audit.
- Exposed dual execution modes — direct-by-ID and semantic lookup-by-client + bot — with a
request_idthreaded through the platform's central logging pipeline. - Tuned the connection pool for production (pre-ping, hourly recycle, strict 10-connection cap) with explicit pool-exhaustion 503 handling, shipping fire-and-forget structured execution logs to a shared logging service.