Work
Staff Software Engineer with 16+ years across AI agent platforms, frontend, and backend. Founded ServiceNow's Agentic Developer Platform, led a 9-person team, and ships agent systems end-to-end: tool orchestration, eval-gated rollout, governance, and full-stack React/Java ownership.
What I do
Agentic AI & MCP
Designed and shipped MCP-driven agentic workflows integrated into the engineering lifecycle, adopted by 20+ teams with roughly 95% less research and debugging effort.
Local LLM inference
Run MLX-optimized models (Qwen, Gemma, DeepSeek) locally on Apple silicon via Ollama, plus memory-efficient loaders like turbo-fieldfare that stream model portions on demand: Gemma 4 26B under 2GB RAM. Took VLMs end-to-end on-device: Qwen2-VL-2B on llama.cpp/GGUF via Metal across a quantization ladder on iPhone, with sub-second TTFT from architecture and resolution levers.
Developer platforms
Founded the Agentic Developer Platform: multi-team bootstrap and governance infrastructure for agents, skills, and MCP servers.
Frontend architecture
Architected scalable React and TypeScript systems powering low-code tools like Table Builder and Workspace Builder.
Backend & data ownership
Hands-on backend behind the surfaces: Python/FastAPI and SQLite services with guarded, read-only SQL access, plus Java/Glide enterprise case backends and REST APIs.
System design
Metadata-aware, multi-model architectures and design patterns built to hold up across teams, tenants, and releases.
Quality & reliability
Enterprise testing strategy with Jest, React Testing Library, and Playwright; performance optimization and WCAG/ARIA accessibility. Lifted legacy coverage from 20% to 90%.
Technical leadership
Lead a 9-person engineering team, run technical hiring and high-bar architectural reviews, and mentor engineers in AI-first development practices.
Agentic Developer Platform
Multi-agent orchestration. 20+ skills, 20+ agents, 20+ teams, four weeks from concept to production.
Selected work
Agentic Developer Platform (ADP)
Multi-team bootstrap and governance infrastructure for Claude Code, MCP servers, agents, and skills. Built in four weeks on a multi-agent orchestration pattern, rolled out to 7 pilot squads across 20+ teams with 20+ skills and 20+ agents.
✓ 20% → 90% legacy coverage
✓ open defects → single digits org-wide
✓ Days → hours
✓ Multi-agent pattern
MCP-driven engineering workflows
Integrated MCP-driven agentic workflows into the engineering lifecycle organization-wide. Reduced research and debugging effort by approximately 95% while improving accuracy.
✓ 20+ teams
✓ Higher accuracy
Build Agent, ServiceNow Studio
Elevated App Engine with agentic workflows for AI-assisted application creation, flow generation, and tool orchestration across 35+ metadata types and 11 domains. Built on the planning pattern (think first, then execute) with a metadata-aware, multi-model (Claude default; Now LLM + Claude on Bedrock), self-healing agentic architecture.
✓ Multi-model
✓ Self-healing
✓ Metadata-aware
✓ Planning pattern (think → execute)
Release lifecycle & app summary agents
Automated delta analysis, human-readable release notes, and CAB-ready deployment manifests: documentation time from hours to a single click. Built on the REAct agentic framework (reason → act → observe). Ships as the Release lifecycle documentation agent in Now Assist for Creator.
✓ CAB-ready manifests
✓ Delta analysis
✓ REAct framework
Results
Every number below is sourced from the work it describes.
On-device, on-desk
# desk: Apple silicon $ ollama run qwen3 && turbo-fieldfare load gemma-4-26B ✓ 26B under 2GB RAM (streamed on demand) # phone: iPhone A16, Qwen2-VL-2B, GGUF ladder, 270 runs $ ordo bench --ladder q4_k_m..q2_k --device a16 ✓ TTFT 30s to 805ms (architecture and resolution levers)
Everything shipped
Every live build, with what each one changed. The home page shows a hand-picked four.
SpendIQ Copilot
Agent · on home14-case golden-set eval harness passes with majority voting: agent answers enterprise data questions without inventing numbers
Conversational spend-analysis agent: LangGraph ReAct over SQLite + FTS5 with SQL guardrails (SELECT-only, read-only connections), an AST-whitelisted calculator instead of eval, and a 14-case golden-set eval harness with majority voting, all passing
ordo
iOS app · on homeTime-to-first-token cut from 30s to 805ms on an iPhone A16 by changing architecture and resolution, not hardware
On-device vision-language model built on Qwen2-VL-2B with llama.cpp across a GGUF quantization ladder, deployed to iPhone (A16), with a private photo eval set, per-stage TTFT decomposition, and LoRA accuracy recovery
Revision
Desktop app · on homePrincipal-level interview prep that works fully offline from a single SQLite file: no account, no sync, no network
Local-first Tauri desktop app for principal-level interview prep: FSRS-5 spaced repetition, drag & air gestures, single SQLite file, fully offline. DSA / system design / AI / behavioral decks with glassmorphic review, cloze + KaTeX, and 53-week analytics
Switchyard
Simulator · on homeEvery internal step of an agentic RAG loop (ingestion, retrieval, tool calls, memory) is watchable instead of imagined
Interactive simulator for an agentic AI + RAG pipeline: watch ingestion, retrieval, tool calls, memory, and a live trace of every internal step while the loop runs.
jev-browser
LibraryOne observed element and one operation per browser step, with zero runtime dependencies and no screenshots in the loop
Harness-agnostic TypeScript core that puts Jev (TypeSafe System One) in charge of browser decisions: one observed element and one operation per step, text-only payloads with no screenshots in the loop, zero runtime dependencies
job-search-skills
Agent skillsA whole job search (sweep, rank, autofill, submit) runs locally against the real eligibility rules
Two portable agent skills that run an entire job search locally: job-finder sweeps boards and ranks roles by real eligibility, apply-to-jobs autofills and submits ATS applications in your own browser, and a Jev decision layer keeps uncertain answers uncertain
intent
ProtocolAgent-driven changes ship with declared scope, protected tests and a green baseline enforced by a verifier
Repo-owned protocol and verifier for agent-driven changes: declared scope, protected tests, digest-bound approvals, and a green baseline
job-radar
Data pipelineSenior+ India engineering roles land on a static site with JSON/RSS feeds, with zero runtime dependencies
Daily crawl of senior+ frontend and AI engineering roles in India from company ATS boards into a static site with JSON/RSS feeds; zero runtime dependencies, forkable configuration
tutor
Agent skillTutor state on disk survives a clean context, so difficulty tuning carries across sessions
Adaptive tutor skill for AI coding agents: one question per turn, difficulty tuned to the learner, state on disk that survives a clean context
Experience
ServiceNow
- Designed ALA's agent architecture (tool contracts, delta-analysis flow); owned requirements → EA Mar 2026 → GA Apr 2026.
- Set the App Summary consolidation architecture: Now Assist skills → AI Agent Studio; one generic flow replaces 4-provider settings (Now LLM, Gemini, OpenAI, Anthropic).
- Drove ADP adoption: legacy test coverage <20% → 95% in 4 weeks across update sets, source control, and team dev; 7 pilot squads.
- Stabilized Table Builder and Workspace Builder; grew Table Builder entry points 1→4, owned Studio support, UI/API fixes, and frontend performance.
- Instrumented agent quality in production (reasoning traces, tool-selection/hallucination tracking, cost/latency dashboards feeding evals) and shipped human-gated plan → approve → execute → verify with rollback.
- Tech-lead 9 engineers: ~60% hands-on coding, PR/architecture reviews, mentoring, cross-team coordination; Java/Glide full-stack ownership.
Rippling
- Contributed to frontend infrastructure and global engineering initiatives; drove codebase improvements, interviewed candidates, and mentored junior engineers.
Reputation
- Led the Actions Team frontend, migrating Angular → React and cutting page load time in half.
- Delivered the Actions feature and management dashboard; raised accessibility compliance from 80% → 95%.
Worktech Software
- Built products and services with JavaScript, Node.js, MongoDB, Meteor, and REST APIs.
- flexzone.in (2018–2020): café aggregator for coworking spaces; hospeto.in (2017–2018): doctor video consultations with records and billing.
Eze Software Group
- Built an order-management system for real-time short-selling, plus EzeLocate and DataFeed for short inventory and real-time stock data (REST APIs, MS SQL).
Oracle
- Built Plan Copy and Benefits Relationships systems (Java, Oracle Database, SQL/PLSQL); cut HRMS test-to-production data setup by 80%.
Skills
AI & Agentic Engineering
AI agents, Agent UIs, MCP workflows, Claude Code, LLM APIs (Anthropic, OpenAI, Gemini), LangChain / LangGraph, Prompt engineering, Tool orchestration, Multi-model systems, Multi-agent orchestration, REAct agents, LLM evals (golden sets, LLM-as-judge), Agent evaluation & guardrails, Human-in-the-loop agent design (plan → approve → execute → verify), AI governance (AI Agent Studio / Control Tower), RAG / keyword retrieval (BM25, FTS5)
Frontend & Platform
JavaScript (ES6+), TypeScript, React, Next.js, Redux Toolkit, RTK Query, Node.js, Webpack, Vite, PWA
Backend & Data
Python, FastAPI, SQLite, MySQL, Java, REST / GraphQL APIs, MongoDB
Architecture
System design, Scalable frontend architecture, Design systems & component libraries, Developer productivity, Metadata-driven development
Reliability & Quality
Jest, React Testing Library, Playwright, Performance optimization, Performance profiling & tuning (production regressions, budgets), WCAG / ARIA accessibility
CI/CD & Tooling
GitHub Actions, Jenkins, Docker
Observability
Sentry, LogRocket, Datadog, Sumo Logic, Agent observability (reasoning traces, tool-selection errors, cost/latency), On-call & incident command
Education
National Institute of Technology (NIT), Tiruchirappalli
Master of Computer Applications (MCA), 2007-2010
Banaras Hindu University (BHU), Varanasi
BSc (Hons), Mathematics, 2003-2006
Patent US20150046358
Importing tested objects into benefits programs deployed on production systems. Assigned to Oracle Corp., status Abandoned. Google Patents