Hi, I’m @Harsha03,

I build AI systems
that automate your workflows
and ship as real products.

• iOSClaude Haiku 4.5

  l8ter — Multimodal AI Capture

  Native iOS app using Claude as the core intelligence layer. Classifies image + caption into 9 structured categories via JSON-schema-enforced extraction, then enriches metadata through a separate web-search pass.

  SwiftSwiftUISwiftDataClaude APIMultimodal
  ★ — · ⑂ —
• 4-Agent PipelinePython

  deep_research — Multi-Agent System

  Planner → Search → Writer → Email. Real-time streamed status through Gradio with per-agent error isolation. Each agent exposes a strict Pydantic contract, making components swappable.

  PythonGPT-4o-miniGradioasyncioSendGrid
  ★ — · ⑂ —
• SaaSMulti-user

  Prelegal — AI Legal Document SaaS

  Drafts 12 legal document types (NDA, SaaS, DPA, BAA) through an AI chat that walks users field by field with live markdown preview. Template-suggestion flow routes plain-language descriptions to the closest match. Delivered in one week.

  Next.js 16FastAPIPythonAI Chat
  Read the case study →
• EnterpriseActive

  Multi-Agent Business Automation

  Orchestrates AI agents across HR, IT, Finance, and Customer Operations workflows. Dynamic task assignment with RAG knowledge layer and production-ready governance.

  PythonLangGraphCrewAIAzure OpenAIKubernetesDocker
• EnterpriseShipped

  AI Contract Review Assistant

  Accelerates legal agreement analysis using LLMs and enterprise knowledge. Azure AI Document Intelligence pipeline over complex agreements with structured extraction.

  PythonAzure OpenAILangChainVector DBDocker
• EnterpriseShipped

  Sales Intelligence AI Copilot

  Prepares sales teams for customer engagements using CRM data and LLMs. Pulls context from Salesforce/Dynamics and Microsoft Graph API to generate pre-meeting briefs, competitive insights, and recommended talking points.

  PythonAzure OpenAILangChainVector DBSalesforceMicrosoft Graph API
• EnterpriseActive

  Intelligent Invoice & Document Processor

  Eliminates manual data entry via OCR and LLM pipeline for invoices and receipts. Handles scanned PDFs and photos, validates against business rules, and routes ambiguous cases to a human-in-the-loop review queue.

  PythonAzure Document IntelligenceOpenAILangChainREST APIsDocker
• EnterpriseActive

  Employee Onboarding & HR Knowledge Assistant

  Conversational AI answering HR policy questions via RAG over company documents. Integrates with Slack and Teams, escalates unanswered questions with full context to human agents.

  PythonClaude APIOpenAIRAGVector DBSlackTeams
• EnterpriseActive

  AI Business Intelligence Analyst

  Allows non-technical users to query business data in plain English. Text-to-SQL engine generates charts and narratives with role-based access control, turning complex database queries into actionable insights.

  PythonOpenAIText-to-SQLReactPostgreSQLBigQuery
• RAGTool Use

  Personal AI Agent & RAG Portfolio

  Portfolio chatbot with LangChain + Chroma + OpenAI embeddings, function-calling lead capture and Pushover notifications. Plus a Kanban PM with full-stack JWT auth and AI sidebar assistant (44 Vitest + 71 pytest tests + Playwright E2E).

  LangChainChromaOpenAIPythonVitestpytest

Sharing

I write about the AI patterns I use in production — RAG optimization, agentic architectures, and what it takes to ship LLM systems that actually hold up.

• LinkedInEnsemble AI is reshaping fraud detection — pairing predictive ML with encoder LLMs creates a system that is both faster and smarter.2 likes
• LinkedInContext engineering vs prompt engineering — why assembling everything the model needs at runtime matters more than shaping tone and format.3 likes · 1 comment
• LinkedInDeep Agents: solving the long-horizon challenge. How planning tools, sub-agents, and file system memory keep agents on track for hours.3 likes
• LinkedInThe case for cost-optimized RAG — smarter models, tuned chunk sizes, and tiered storage to keep production costs in check.2 likes · 1 comment
• LinkedInLLM security: testing isn't optional. Prompt injections, jailbreaks, and model poisoning — and how to mitigate each.2 likes
• LinkedInParallel is reimagining web search — not for humans, but for AI. Its Deep Research API turns hours of manual investigation into minutes.4 likes
• LinkedInLow-latency RAG: most delays come from LLM inference, not retrieval. Router LLMs, caching, and quantized models can fix that.2 likes
• LinkedInMoving RAG from proof-of-concept to production brings new hurdles — performance, latency, data quality, and security vulnerabilities.3 likes
• LinkedInMultimodal RAG: most systems stop at text, but enterprise knowledge lives in PDFs, slide decks, images, and audio.3 likes
• LinkedInLLM observability in RAG: you can't fix what you can't see. Trace every step — retrieval, reranking, prompt construction, generation.1 like
• LinkedInMulti-agent orchestration: single-agent LLMs are hitting their limits. Protocols like A2A and ACP standardise how agents collaborate and delegate.3 likes
• LinkedInFrom deterministic code to adaptive intelligence — LLM-native apps interpret intent and adapt in real time instead of coding every edge case.3 likes

Education

Experience

From biomedical hardware to production AI systems — I move fast across domains and ship things that work.

• Lu
  LumiviciAug 2024 — May 2025 · Tucson, AZ

  Design Engineer Intern

  Led development of a self-sterilising far-UVC face shield. Used LightTools optical simulation to raise interior UV coverage from 72% to 95%+, then owned V&V of the full technical data package.

  • ›Raised interior UV coverage from 72% to 95%+ through iterative LightTools simulation.
  • ›Achieved ≥4-log (99.99%) viral reduction in clinical airflow tests at the University of Arizona WEST Center.
  • ›Owned the complete verification & validation package for FDA submission pathway.
• Gu
  Gutruf Lab, University of ArizonaJun 2021 — Apr 2022 · Tucson, AZ

  Research Intern

  Fabricated flexible, battery-free spinal-cord stimulators for in-vivo neuromodulation studies.

  • ›Fabricated polyimide PCB-based devices with Parylene-C encapsulation for biocompatibility.
  • ›Validated devices for in-vivo studies in collaboration with neuroscience research team.

Competencies

Let's build something real.

Available full-time in San Francisco. I respond fast and give honest assessments of what's feasible. Send a short message — even if it's vague.