Yulin Deng

Global AI Startup | AI Application Developer / AI Platform Backend Engineer

Jan 2025 - Present

• Led the migration of an enterprise RAG service from LangChain to LlamaIndex, reorganizing it into layered API, ingestion, parsing, embedding, retrieval, reranking, and scheduling modules while unifying REST and MCP query paths.
• Designed connectors for AWS S3, Google Drive, SharePoint, and SQL Server with full, incremental, scheduled, and event-driven synchronization; used Celery / SQS batching, retries, import deduplication, and status tracking to improve processing efficiency by ~50% and reduce sync latency from hours to minutes.
• Built a multimodal processing pipeline for PDF, Word, PowerPoint, Excel, email, and image content, integrating OCR, layout and table extraction, image processing, and normalized attachment metadata for retrieval and source citation.
• Implemented Weaviate metadata filtering, query rewriting, reranking, and source-level authorization, with model integrations across OpenAI, AWS Bedrock, and Azure OpenAI; reduced query latency by ~30% and runtime memory usage by ~25%.
• Built MCP server and tool-calling capabilities with OAuth and JWT validation, enabling agents to access enterprise documents, Google Workspace, and databases within user authorization boundaries, supported by actionable logs and metrics.

Stack: Python / FastAPI / LlamaIndex / Weaviate / pgvector / Redis / Celery / SQS / OpenAI / AWS Bedrock / Azure OpenAI / MCP

Shanghai JIDU Automotive Co., Ltd. | R&D Productivity Platform / Enterprise Integration Engineer

Jan 2022 - Dec 2024

• Integrated 10+ internal systems, including Jira, Jama, Confluence, Feishu, and feedback, bug, and vehicle-issue platforms; built REST endpoints, Behaviours, Listeners, and automated workflows with ScriptRunner / Groovy, increasing process automation by ~85% across PMO, engineering, and supplier collaboration.
• Designed a Webhook-to-MQ architecture that reduced direct Jira API load and improved scalability and fault isolation for cross-system event processing.
• Led the company-wide domain-account & password platform (onboarding/offboarding, external-vendor accounts, permission-change alerts, Feishu audit logs, MFA), using JWT to integrate Moka, Active Directory, and HR master data across Web / Feishu / mobile / WeCom; consumed ticket messages via MQ to maintain security-group access across systems, cutting onboarding/offboarding from days to hours and saving more than RMB 2 million in annual license costs.
• Optimized the org-structure sync mechanism by breaking up large transactions and eliminating no-op updates, cutting sync-task runtime from an average of 15 minutes to under 3 minutes and significantly reducing AD-server load; diagnosed and fixed production issues such as AD password desync and cross-server latency to keep the system stable.

Stack: Spring Boot / MyBatis / Redis / RocketMQ / Jira / Groovy / ScriptRunner / Elasticsearch / AD / LDAP / Feishu

Shanghai Shurui Information Technology Co., Ltd. | Python Backend Developer / Engineering Standards Lead

Feb 2021 - Jan 2022

• Standardized Django / Flask project scaffolding and deployment practices, introducing consistent layering, coding conventions, code review, and Git Flow to improve delivery consistency and maintainability.
• Refactored 5,000+ lines of core recruitment-system code and resolved 20+ legacy defects; reduced average API response time from 2.5 seconds to under 0.8 seconds through application and query optimization.

Stack: Python / Django / Flask / MySQL / Redis / Git Flow

Shanghai Data Exchange Co., Ltd. | Python Backend Developer

Feb 2020 - Jan 2021

• Developed core business services, APIs, and cross-system integrations for data-exchange workloads; led async refactoring, module restructuring, and SQL / index tuning to improve throughput and API responsiveness, reducing selected query times by ~50%.

Stack: Python / Django / DRF / Flask / FastAPI / MySQL / PostgreSQL / Redis / RabbitMQ / Celery / Docker / Nginx

Shanghai Yilu Investment Management Co., Ltd. | Backend / Systems Developer

Feb 2018 - Jan 2020

• Built a hotel SaaS PMS and business platform from the ground up, covering orders, payments, membership, catering, and room-status domains, serving both Web and self-service check-in machines, deployed at 200+ hotels. (Technical architecture in Selected Projects below.)
• Contributed to the business-platform architecture, re-architecting the order, payment, resource, user, and membership centers into independently / composably deployable microservices, driving parallel front-/back-end development and multi-client SDK auto-generation with Swagger contract-first design, and abstracting third-party standard interfaces for unified integration to speed up product iteration (Spring microservice stack).

mate-assist | Automotive Sales & Operations Agent Backend

Independent personal project (in parallel with employment), 2026 | FastAPI / LangGraph / MySQL / Redis / Celery / Weaviate / Kubernetes / KEDA

• Designed an 11-node LangGraph conversation graph: context control → multimodal image grounding → query rewriting (coreference resolution) → intent classification with deterministic table-based routing → retrieval + relevance grading + low-quality re-write & re-retrieval (CRAG) → sales-script generation → long-term memory extraction, executed across knowledge, sales-support, discourse, and vision branches.
• Strictly separated conversation ground truth (messages) from the per-turn model context (model_input), writing retrieval/summary/memory injections only to the temporary context to eliminate history pollution at the root; persisted last_intent / last_car across turns to fix intent drift where multi-turn follow-ups were re-judged statelessly.
• Compressed history with incremental rolling summaries (old messages folded into a summary and trimmed out), keeping per-turn token cost near-constant for long conversations; MySQL checkpointer / store power cross-session memory, conversation replay, answer regeneration, and restart recovery.
• Delivered SSE streaming + TTFT / end-to-end latency metrics + node-level tracing + error-type-aware retries (deterministic errors like content_filter are not retried, only transient ones); diagnosed and fixed a media-pipeline performance issue, cutting one media Q&A from ~2m07s to a normal range.
• Built a compliance hard-gate for sales scripts (red-line blocklist + service-request fallback scenes) and a full end-to-end black-box regression suite (intent continuity / memory bleed / compliance / vision / media), with run_all_regression as a CI gate; every production bad case is hardened into a regression test.
• Built Celery offline batch jobs (customer profiling / sales-funnel / store briefings / lost-lead analysis) with exponential backoff, concurrency gates, cache short-circuiting, and KEDA autoscaling from 0–10 workers by Redis queue depth; raising the concurrency gate from 8 to 20 cut gate-wait P50 from 32.5s to 10.5s.

Jarvis Registry | Enterprise MCP / A2A Agent Gateway

Core contributor | 2.8k+ GitHub stars | Nearly 300 authored commits on main | Python / FastAPI / MCP / A2A / OAuth 2.0 / FAISS / OpenTelemetry

• Contributed across the enterprise MCP / A2A gateway: server & agent registration, FAISS-based semantic tool discovery (vector search over skills / descriptions / tags so agents locate the right tool at runtime instead of stuffing every tool into the prompt), reverse proxying, federation, the Skill Registry, and integrations for Cursor, Claude Desktop, GitHub Copilot, and VS Code.
• Made the gateway stateless by moving client-info into a Redis cache for horizontal scaling; built multi-provider OAuth 2.0 / OIDC across Keycloak, Amazon Cognito, Microsoft Entra ID, and Google, with JWT / session validation, scope-based RBAC, object-level ACLs, and tool-level access control.
• Contributed to MCP / A2A workflow orchestration (workflow-worker executor / scheduler), scheduled execution, node-level failure attribution and recovery, plus structured logging, OpenTelemetry tracing, Prometheus metrics, Tempo / Grafana integration, and sensitive-trace controls.

rag_api | ID-based RAG Indexing / Retrieval Microservice

Personal project | LangChain / FastAPI / PostgreSQL + pgvector / asyncpg

• Organized embeddings by file_id for file-level targeted retrieval backed by document metadata, giving natural multi-tenant / permission isolation and per-file incremental rebuilds (delete / update at file granularity, no full-corpus re-index).
• Abstracted the embedding provider to be pluggable (OpenAI / Azure / Bedrock / HuggingFace / TEI / Vertex AI / Ollama) with pgvector and Atlas MongoDB vector stores interchangeable, decoupling model and dimensionality.
• Built a multi-format document parsing pipeline (PDF / Word / PPT / Excel / CSV / Markdown) with RapidOCR for extracting image content from PDFs.
• Fully asynchronous: an asyncpg connection pool plus a bounded shared thread pool for CPU-intensive parsing / embedding so the event loop never blocks; pgvector indexing and query optimization for high-concurrency ingestion and retrieval.

yilu_ispider | Hotel SaaS PMS & Self-Service Machine Backend

Hotel business platform, 2018 - 2020 | Django / DRF / Celery / RabbitMQ / Nameko / Redis / MongoDB / Docker

• Built the hotel PMS backend from the ground up on Django / DRF, organized into modular apps by domain (orders, payments, membership, catering, room status / pricing, permissions), with JWT auth at the API layer, serving both Web and self-service check-in machines.
• Handled payment verification and order timeouts with RabbitMQ delay + dead-letter queues for async decoupling and timeout recovery, paired with Celery + scheduled jobs (beat) for room-status refresh, reconciliation, and batch processing.
• Delivered the self-service machine’s room lock / unlock interfaces wired to order cancellation, keeping room status and inventory consistent under multi-client concurrency; provided room-status maps and check-in operational APIs.
• Introduced Nameko for inter-service RPC, Redis (with Sentinel HA) caching + django-haystack search for read performance, and MongoDB for unstructured business data.
• Split configuration across environments / tenants (dev / pro / pmsb branches and settings) with Docker containerized deployment, isolated per-hotel environments, and gray releases.

Jiangxi University of Science and Technology | B.Eng. in Electronic Information Engineering | 2013 - 2017