Founding Software & AI Engineer at VelocitiPM
About three years of shipping AI features in real products, from enterprise service desks at ServiceNow to the React front end and Python AI service of an early-stage SaaS app.
I'm Devendra Sai Mupparaju, a founding software and AI engineer at VelocitiPM, where I work on the React and TypeScript front end and the Python service behind its AI ticket generation. Before that I rolled out generative AI incident summarization at ServiceNow, and before that I built data pipelines at Cognizant. In each job the AI worked only as well as the data going into it and the review step after it, so I build both.
At VelocitiPM I build the React and TypeScript screens product teams plan their sprints in, and the Velo AI interface that drafts tickets and user stories. Behind it, I feed product design context into the Python AI service so the drafts match what was actually designed. Users edit every draft before it is saved.
At ServiceNow I configured Now Assist to summarize incidents and draft resolution notes and knowledge articles. I built the Flow Designer actions that sent incident details to the AI service and routed drafts to agents for review, and wrote Python checks that compared each response with the source incident.
At Cognizant I built reusable Python and SQL pipelines that turned raw enterprise data into datasets for analytics and machine learning. I cleaned it with Pandas and NumPy and added validation checks that caught bad records before they reached downstream systems.
I use Claude, Codex, and Antigravity most days to read unfamiliar codebases, prototype, debug, and refactor. They speed the work up and sometimes catch what I miss. The architecture and validation calls stay with me.
I started on the data side at Cognizant, then put generative AI in front of service desk agents at ServiceNow. My MS research at Iowa State took me inside the models: with activation steering I raised CodeLlama-7B's HumanEval accuracy by 8 points without retraining it. At VelocitiPM I use all three on one product, and I would rather spend the extra day getting a feature right than ship something that only works in a demo.
Product planning and sprint management platform. I build the React and TypeScript front end and the Python AI service behind its ticket generation.
Generative AI for IT service desks on the ServiceNow platform: incident summaries, resolution notes, and knowledge articles that agents review before use.
Data pipelines and preprocessing for enterprise analytics and machine learning work.
A full-stack grocery e-commerce platform that deploys to GitHub Pages through an automated CI/CD pipeline. Next.js SSG with generateStaticParams pre-renders every product route at build time, which got rid of the runtime 404s as the catalog grew. A fallback layer switches between live Supabase/PostgreSQL and local mock data, so deploys across environments never go down.
A 'Time Capsule' feature built on Supabase, which pushes video updates instantly and gave me a way to test how the real-time data scaled. The TypeScript dashboards update as data comes in and make complex hobby progress readable at a glance.
5-agent pipeline that automates the entire job application lifecycle: Agent 1 scouts jobs from LinkedIn/Indeed via Apify with relevance scoring and deduplication. Agent 2 scrapes ATS forms (Greenhouse/Lever) and maps every field with Gemini. Agent 3 generates tailored cover letters and elevator pitches per role. Agent 4 answers every unique form question from the candidate profile. Agent 5 fills and submits via Playwright browser automation, pausing for human confirmation before the final submit. Running alongside the pipeline are a Gmail IMAP watcher, a resume tailor with ATS scoring and PDF generation, and a real-time React + Firebase Kanban board.
Autonomous content pipeline that pulls trending AI topics via Perplexity API, generates captions with Gemini 2.5-Flash, creates images using Imagen 3, and publishes straight to Instagram via instagrapi. The whole thing runs on a schedule with no manual steps and no paid publishing APIs. Image generation falls back through Imagen 3 → Pollinations → Pillow, and a custom slide compositor handles the topographic texture and gradient typography. Four external APIs are wired in with retry logic and structured CSV logging.
Web navigation agent that takes voice or text commands and works through any website on its own, using AI vision and browser automation. Set-of-Mark element tagging lets Gemini visually parse and click UI elements, so no site-specific scripting is required and it works on any site. Voice input is supported throughout.
A local AI agent turned into a personal assistant across four connected builds. It runs a free OpenRouter model on Windows and stays reachable from an iPhone anywhere over a private Tailscale mesh. Wired to Gmail and Google Calendar through OAuth2, it reads, drafts, and sends on command, always pausing for confirmation before anything goes out. A custom SKILL.md taught it to compile a weekly digest of GitHub commits and calendar events and deliver it to the phone automatically every Friday via cron — the difference between raw API access and an agent that actually formats useful output.
An English speaking coach that runs entirely in the browser, no install. Pick casual practice or a mock interview for a specific role, speak through the mic, and hear the AI reply aloud. Conversation state lives in Gemini's Interactions API server-side, so there's no growing history to forward on every turn. Each session ends with a scored report — grammar, vocabulary, clarity, relevance — returned as structured JSON through a schema attached to the request, with no string parsing on the server. Voice in and out runs through the Web Speech API with no paid speech service.
Formal verification of fairness properties in a 3-client federated learning protocol. Models the FL workflow as a finite-state machine in NuSMV, encodes CTL properties for liveness, safety, and fairness, then automatically verifies no client is starved and no deadlocks occur across bounded aggregation rounds.
A 2D arcade physics engine tuned to run fast in mobile browsers. 'Cushion' collision detection keeps it smooth at 60 FPS. The daily challenge is deterministic: seeded randomization gives thousands of concurrent players identical physics conditions.
Focus on Artificial Intelligence, Machine Learning, and System Architecture. GPA: 3.5/4.0
Focus on Software Engineering, Data Structures, and Algorithms. GPA: 7.8/10
Send a message and I will get back to you.
"Hi. Ask me anything about Devendra's work."
