Darshan Vaghasiya

AI Engineer · Software Engineer

darshangvaghasiya0@gmail.comgithub.com/Dars-gitLinkedInOpen to new roles

Software engineer who builds AI products end to end: Rust services, Python LLM agent runtimes, and TypeScript interfaces. At Dialogica I shipped an AI legal assistant for law firms that ran in production with pilot users at real firms, covering human-in-the-loop approvals, a 150+ tool agent registry, and document, email, and calendar integrations. Earlier work spans distributed microservices, LLM inference pipelines, and cloud MLOps. I use AI coding agents and LLM tooling daily, which shapes how I design tool interfaces and approval flows for the agents I build.

Experience

Dialogica

Mar 2026 – Sep 2026

AI Engineer · Santa Monica, CA

  • Built a human-in-the-loop approval system in Rust and SolidJS gating AI email, document, and calendar actions.
  • Designed a tool-provider registry unifying 150+ MCP tools across the Rust backend and Python LangGraph runtime.
  • Extended the MCP tool surface as integrations grew, keeping schemas and error shapes consistent across providers.
  • Implemented pause/resume and reconnect-safe SSE streaming so approvals and tool results survived dropped connections.
  • Integrated iManage DMS and SharePoint over OAuth 2.0, clearing a 71-finding audit against a Microsoft 365 tenant.
  • Hardened the LangGraph runtime with argument validation, tool-result truncation, and error recovery for pilot use.
  • Traced LangGraph runs in LangSmith, using tool-call and token traces to find failures before pilot users hit them.

Blue Sigma Software Solutions

Jul 2025 – Feb 2026

Software Engineer · Baltimore, MD

  • Engineered Java and Spring Boot microservices on AWS, with Kafka carrying events between distributed services.
  • Secured REST APIs with OAuth 2.0 and JWT, standardizing authentication and meeting compliance requirements.
  • Orchestrated CI/CD on Jenkins, Docker, and Kubernetes, moving releases to rolling deploys with no downtime.
  • Instrumented Redis caching and Grafana dashboards, using load tests to catch latency regressions pre-release.
  • Authored runbooks and design docs for team-owned services, shortening on-call triage for unfamiliar incidents.

VITG

Nov 2024 – Jul 2025

Software Engineer · Ellicott City, MD

  • Shipped AWS Lambda ETL workflows enriched with LLM inference, cutting manual reporting effort by 35%.
  • Wired Claude multimodal OCR pipelines into document intake, lowering human review time by 30%.
  • Delivered an LLM chatbot joining geospatial data with natural language queries for city-planning teams.
  • Tuned LLM inference pipelines for cost per token, throughput, and reliability across multimodal workloads.

University of Maryland, Baltimore County

Jun 2023 – Jul 2025

Research Assistant (Academic) · Baltimore, MD

  • Rewrote malware scanning modules in C++ with multithreading and memory tuning, raising detection speed by 20%.
  • Improved Python and C++ scanner components for Linux, raising detection accuracy and runtime stability.
  • Simulated cyberattacks in Dockerized test environments to stress ML-driven detection under realistic load.

Projects

Darshan Marketplace: Cloud-Native E-commerce

github.com/Dars-git/Fullstack-E-commerce-web-application

Java 21 · Spring Boot · Spring Cloud Gateway · Eureka · MongoDB · React · Vite · Docker · Kubernetes · Helm · Terraform · AWS EKS

  • Built a React (Vite) storefront with auth, product search, categories, and cart on top of Spring Boot microservices.
  • Split the backend into gateway, registry, auth, user, product, category, cart, order, and notification services, each with its own MongoDB database.
  • Provisioned AWS with Terraform (VPC, EKS, autoscaler, ALB controller, ECR) and shipped every service with its own Helm chart.

Personalized Video Recommendation Engine

Python · TensorFlow · BigQuery · Dataflow · Kubernetes · Docker · Faiss

  • Trained ranking models over video and text features in TensorFlow, scored against held-out watch data.
  • Served the model on Kubernetes behind a REST API, holding sub-100ms P95 latency under local load tests.
  • Indexed embeddings in Faiss for approximate nearest-neighbour retrieval in under 50ms per query.

AI Video Summarization Platform

Python · PyTorch · Transformers · GCP (GKE, Cloud Storage) · TensorRT

  • Fine-tuned a pretrained summarization model on a 100K-sample dataset, evaluated with ROUGE-L scoring.
  • Quantized inference to INT8 with TensorRT, cutting benchmark latency and GPU cost by about 40%.
  • Compared generated summaries against reference text to tune decoding and prompt settings.

Ask-Your-Docs: AI Document Q&A

github.com/Dars-git/Ask-Your-Docs

Python · FastAPI · LangChain · Pinecone · Streamlit

  • Assembled a document Q&A service on FastAPI and LangChain with Pinecone namespaces per document set.
  • Routed queries across OpenAI and Anthropic models with tunable temperature and retrieval depth.
  • Scored retrieval against a 30-question eval set, tuning chunk size and top-k against hit rate.

Go · gRPC · Protocol Buffers · YAML

  • Replicated a token store across three gRPC servers, committing writes only after a majority of acks.
  • Made reads atomic with read-impose-write-all: return the newest timestamp only after writing it back to a quorum.
  • Simulated fail-silent replicas (sleeps, negative acks) and kept the system correct with one node silent.

SmartRoutePlanner: AI Travel Planner

github.com/Dars-git/SmartRoutePlanner

Next.js · TypeScript · Convex · Clerk · OpenAI · Google Maps · Stripe

  • Generated trip plans in three parallel LLM batches with JSON-schema function calling, each wrapped in scheduled retries.
  • Streamed each section to the browser as it landed through Convex reactive queries.
  • Added shared plans with rate-limited email invites, expense tracking, live weather, and payments.

Node.js · TypeScript · Express · Kafka · Kafka Streams · Redis

  • Validated device alerts at an Express API and published them to a partitioned Kafka topic.
  • Enriched alerts in a Kafka Streams worker with recipients looked up in Redis per builder and building.
  • Routed invalid or unroutable alerts to a dead-letter topic with a reason, so bad input never blocks good alerts.

Java · Spring Boot · Spring Cloud · OAuth2 · MongoDB · RabbitMQ · Docker Compose

  • Personal-finance backend built on the PiggyMetrics reference architecture: accounts, statistics, and notification services.
  • Spring Cloud Config, Eureka discovery, and an API gateway, with an OAuth2 auth server securing service-to-service calls.
  • RabbitMQ (Spring Cloud Bus) for config refresh, Hystrix metrics dashboards, and one-command startup with Docker Compose.

Skills

Languages
Rust, Python, TypeScript, JavaScript, C++, Java, SQL
AI & Agents
MCP servers and tools, LangGraph, LangChain, Tool calling, Human-in-the-loop design, RAG, Prompt engineering
LLM Platform
Anthropic and OpenAI APIs, SSE streaming, Tracing and observability
Backend & Systems
Axum, FastAPI, Spring Boot, Node.js, REST APIs, OAuth 2.0, JWT, Kafka, Redis, asyncio
Data & Storage
PostgreSQL, Pinecone, MongoDB, SQLite, BigQuery, Faiss, ETL pipelines, Event streaming
Front-End & Desktop
SolidJS, React, Tauri, HTML, CSS
Cloud & MLOps
AWS, GCP, Docker, Kubernetes, CI/CD, Jenkins, Grafana, Model deployment, Cloud monitoring
Testing & Practices
System design, Data structures, API versioning, Load testing, Agile

Education

University of Maryland, Baltimore County · M.S. in Software Engineering
Jan 2023 – May 2024
Gujarat Technological University · B.E. in Computer Engineering
Jul 2018 – May 2022