I work on LLM systems in production — the agent loop, the prompt layer, the APIs and datastores under them. Mostly the parts that make language models dependable: planning, tool orchestration, memory, structured output, latency, and evaluation.
AI Engineer at Strategy (formerly MicroStrategy) · Pune, India
🌐 varunmotiyani.xyz · portfolio, projects, and writing
-
AutoAgents — the agent framework behind Strategy's AI products, where a planner turns a complex question into a tool plan and runs it. I've built major pieces of it: the ReWOO/ReAct planning loop, a one-point tool-integration layer so new tools plug in without touching the planner, parallel tool execution for latency, the MCP client/host layer, and SimpleMem — a short- and long-term agent-memory package (ClickHouse + Redis) I designed and own, reused across several internal AI services.
-
AutoDash — turns a natural-language request into a built analytics dashboard. I contributed the ReAct orchestration engine over parallel intent agents, run as a two-stage async task with a fallback for complex sequential work.
-
AutoPrompt — a microservice I built for provider-aware prompt management: prompt versioning and dynamic rendering keyed on model and provider, the piece that makes bring-your-own-LLM work across products.
-
Support-case routing — an LLM service that reads each incoming support case, infers the product area, complexity, and skills involved, scores the available engineers by current workload, and auto-assigns on a short polling loop. Built end to end in async Python — FastAPI, PostgreSQL, Redis.
-
Bootcamp grading — an LLM grader for sales-training submissions: it pulls the trainee's pitch from a call transcript, scores it against a rubric, folds in quiz results, and returns a final grade with a written breakdown. Schema-constrained extraction; FastAPI + React.
-
EP review automation — automates the enterprise performance-review cycle: a parallel pipeline that scores manager feedback against competency frameworks, with video/audio/document ingestion (Whisper, MoviePy), translation, generated DOCX reports, and Teams notifications.
fitness-tracker-ios— an adaptive strength & physique coach for iOS: equipment-aware weekly planning, rolling adaptation around missed sessions and daily readiness, in-gym logging with one-tap machine swaps, and monthly InBody scan analysis. Phone-only, no backend, bring-your-own AI key.dev-playbook— a Claude Code plugin that packages one opinionated method for taking a feature from raw idea to reviewed, merged code: a coordinating session plus fresh subagents per task, with human gates at design, spec, and integration.llmcontrol— one interface for prompts, parameters, and invocation across LLM providers.crawller— a small Python web crawler.
LLM — agents (ReWOO, ReAct, MCP) · prompt engineering · RAG & GraphRAG · knowledge graphs · evaluation (LangSmith, Opik, DeepEval)
- Anthropic Claude Certified Architect – Foundational (CCAF)
- Oracle Cloud Infrastructure — Generative AI Certified Professional
- Neo4j Certified Professional · Neo4j Certified Data Scientist
- Snowflake — Data Warehousing, Data Engineering, and Marketplace badges
Pune, India — open to conversations about AI platform / LLM infrastructure roles.