Ashish Reddy Jaddu
Software Engineer, AI — shipping production AI systems
Backend · AI/LLM · Full-Stack — RAG, Azure, Vertex AI
Software Engineer, AI at Klue, building AI for competitive intelligence. Previously Founding Engineer at LegiSimple, where I shipped 4 product pivots in 20 months — RAG-powered legal research over 120K+ cases, LLM retrieval pipelines, and full-stack workflow automation on Azure, from Python/Django APIs to React Native apps.
Products shipped
4 in 20 months
From GPT legal research → workflow automation
RAG accuracy
82% → 87%
Hybrid chunking + Azure AI Search + Voyage AI
Workflow time saved
3–4 hr → 20–30 min
HR sanctions process for enterprise clients
01 — About
About
Software Engineer, AI at Klue with 3+ years building production RAG systems, LLM pipelines, and AI-powered workflow automation. I architect hybrid search systems with custom relevance metrics, build agentic orchestration pipelines, and ship full-stack products on Azure + GCP that real users rely on.
My journey began in 2020 at HCL Technologies in India, building RESTful APIs and optimizing databases. After earning my Master's in Applied Computer Science from Concordia University (2022-2024), I joined LegiSimple as a founding engineer at the ground floor.
At LegiSimple, I led technical architecture for an AI-native legal-tech startup over 20 months. Shipped 4 product pivots—from GPT-based legal research to a hybrid RAG system over 120K+ case law documents to AI-powered workflow automation—iterating directly with law-firm partners at each stage.
In July 2026, I joined Klue as a Software Engineer, AI — building on the AI platform for competitive intelligence and win-loss analysis, where LLM-powered agents help revenue teams understand why they win and lose deals.
Production RAG System
Built hybrid RAG pipeline over 120K+ legal cases: section-aware parsing, agentic chunking for complex legal sections, semantic chunking for simpler sections, and a custom relevance metric (cosine + citation count + recency + shepherdization). Using Azure AI Search with Voyage AI legal embeddings, ground truth eval improved from 82% → 87% retrieval precision with sub-second latency.
Agentic Orchestration
Designed multi-agent system with an orchestrator coordinating specialized agents: case retrieval, citation validation, and overturned-case detection (shepherdization). Integrated guardrails to validate every citation against the database and enforce safety checks on each tool call.
Workflow Automation Impact
End-to-end HR sanctions platform: Azure Durable Functions orchestrator gathers employee history + handbook + labor regulations, AI recommends action level, human-in-the-loop HR approval, auto-generates compliant letters. 3–4 hours → 20–30 minutes.
Backend Performance & Security
Python/Django APIs optimized with Redis caching + query tuning (30% faster). Azure OpenAI + AI Search in private VNet. Cloudflare WAF protection for sensitive law-firm data. 99.9% uptime.
Cloud Architecture
Design and deploy Azure/GCP infrastructure with VNets, App Services, and serverless functions
AI System Design
Build RAG pipelines, vector databases (Azure AI Search, Pinecone), and LLM orchestration with LangChain
API & Backend
Python/Django & FastAPI REST APIs with caching, rate-limiting, and database optimization for production scale
Security & Data Privacy
Implement VNet isolation, encryption, and compliance controls for sensitive legal data
OpenAI & Azure OpenAI
Production LLM integration, prompt engineering, AI agents
RAG Systems
Vector embeddings, semantic search, retrieval pipelines
Vertex AI
LLM evaluation experiments on a 1K-label ground-truth dataset for retrieval precision measurement and improvement
Vector Databases
Azure AI Search, Pinecone, embeddings storage
LangChain & LangGraph
AI workflow orchestration, multi-step agents
Python/Django & FastAPI
REST APIs, async services, ORM, 30% performance optimization
Azure Functions
Durable Functions, serverless workflows
Node.js/Express
RESTful services, real-time SignalR integration
API Optimization
Caching, query tuning, rate-limiting
Azure
App Services, VNet, Azure SQL, AI Search, private endpoints
Google Cloud
Compute Engine, Cloud Run, Vertex AI
Docker & Kubernetes
Containerization, orchestration, production deployments
Redis & Caching
Query caching, session management, performance
Next.js/React
SSR, Zustand state management, component architecture
TypeScript
Type-safe applications, advanced patterns
React Native
iOS & Android apps published on App Store & Play Store
Real-time UI
Azure SignalR, live dashboard updates
02 — Experience
Experience
Building AI capabilities for Klue's competitive intelligence and win-loss platform — the system revenue teams use to know why they're winning, why they're losing, and what to do about it
Contributing to LLM-powered agent products and insight pipelines that automatically collect, curate, and deliver competitive intel to sellers in real time
Shipped 4 products over 20 months (GPT legal research → Canadian case law → US case law → workflow automation), owning architecture and delivery end-to-end across each stage
Built hybrid RAG pipeline over 120K+ legal cases: section-aware parsing, agentic chunking for complex legal sections (citations, statutes, precedence), semantic chunking for simpler sections, and a custom relevance metric (cosine similarity + citation count + recency + quote level + shepherdization). Using Azure AI Search with Voyage AI legal embeddings, retrieval precision improved from 82% → 87% while keeping query costs sustainable
Designed agentic orchestration system: orchestrator coordinating specialized agents for case retrieval, citation validation, and overturned-case detection. Integrated guardrails to validate every citation against the database and enforce safety checks on each step
Built HR sanctions workflow automation: Azure Durable Functions orchestrator gathers employee history, company handbook, and labor regulations; AI recommends action level; human-in-the-loop HR approval; auto-generates legally compliant letters. Reduced a 3–4 hour manual process to 20–30 minutes
Architected Azure backend: App Services, Durable Functions, Azure SQL, Azure OpenAI + AI Search in private VNet with Cloudflare WAF — 99.9% uptime
Optimized Python/Django APIs with Redis caching and SQL query tuning (30% faster responses, 25% memory reduction); built Next.js/React frontends with SignalR real-time updates and Mixpanel/PostHog analytics
Performed API testing and backend validation across multiple software products, verifying RESTful endpoint behavior, data integrity, and error handling
Conducted AI/ML feature testing for AI-powered products, validating LLM-generated content quality, edge case handling, and response consistency
Executed regression, smoke, integration, and exploratory testing across web applications, mobile apps, and gaming platforms — documented 200+ bugs with detailed reproduction steps
Collaborated with development teams using Jira and Bugzilla to prioritize defects and ensure timely resolution before production releases
Developed and deployed RESTful APIs powering internal React applications; optimized MySQL queries improving page-load performance by ~20% and reducing DB response times by ~10%
Debugged production issues across frontend and backend services; built database-backed defect catalog boosting QA effectiveness by ~40%
Managed deployments for 5 web applications on Heroku, improving configuration and resource usage to cut hosting costs by ~15%
03 — Education
Education
04 — Selected Work
Selected Work
Production systems I've designed and shipped — the problem, the approach, and what it changed.
● Challenge
Law firm partners needed accurate semantic search across 120,000+ NY case law documents. Keyword search missed relevant precedents, and manual research took hours per query.
● Solution
Architected a hybrid RAG pipeline: section-aware document parsing, agentic chunking for complex legal sections (preserving citations, statutes, precedence signals), and semantic chunking for simpler sections. Built a custom relevance metric combining cosine similarity with citation count, recency, quote level, and shepherdization (whether a case is still good law). Added an agentic orchestration layer with specialized agents for case retrieval, citation validation, and overturned-case detection. Guardrails cross-check every citation against the case database and enforce safety checks on each tool call.
✓ Result
85–90% retrieval relevance in production with sub-second latency. Ground truth evaluation improved from 82% to 87% through hybrid chunking refinements, Azure AI Search tuning, and Voyage AI legal embeddings. Actively used in production for legal research.
● Challenge
An enterprise client's HR team handled employee disciplinary cases entirely manually: supervisors wrote warning letters inconsistently, with no compliance checks and no audit trail—creating compliance risk and 3–4 hour processing times.
● Solution
Built an end-to-end workflow platform: supervisors log incidents via a React Native mobile app, triggering an Azure Durable Functions orchestrator. The orchestrator gathers employee incident history, company handbook, and government labor regulations, then uses Azure OpenAI to recommend an action level (verbal warning, written warning, suspension, or termination). HR reviews and approves or overrides. System auto-generates a legally compliant letter and logs every action for audit trails.
✓ Result
Reduced 3–4 hour manual process to 20–30 minutes. Eliminated inconsistent letters through AI-assisted generation. Full audit trail created for every case. Human-in-the-loop design kept HR in control of every final decision.
● Challenge
I wanted to understand production-grade agents beyond frameworks: how to design research and voice agents with full control over planning, tools, observability, and failure modes.
● Solution
Built a Deep Research Agent in raw Python (no LangChain/LangGraph) that plans web research, iterates search→reflect loops, ranks sources with embeddings, and synthesizes long-form cited reports using Azure OpenAI and Tavily. In parallel, built a restaurant voice agent on LiveKit that answers real phone calls, uses Deepgram STT/TTS and Azure OpenAI tools for menus and reservations, and logs all calls and bookings to Supabase.
✓ Result
Demonstrated end-to-end agentic systems with structured logging, retry/backoff, cost controls, evals on a labeled dataset, and real-time voice interactions. These projects are my playground for experimenting with new LLM models and agent patterns outside of production client work.
● Challenge
Field workers needed to capture incident reports and receive case documents on-site. Web-only access created delays and prevented real-time updates in the field.
● Solution
Built cross-platform React Native apps (published on iOS App Store and Google Play Store) integrated with the workflow automation backend. Push notifications for status updates, real-time sync via Azure SignalR, and seamless handoff to the Next.js dashboard for HR reviewers.
✓ Result
Enabled on-site incident capture and real-time document delivery for field workers. Provided the multi-channel access layer that made the HR sanctions workflow end-to-end.
05 — Contact
Get in Touch
Always happy to talk AI systems, RAG, backend engineering — or just to connect. Reach out directly and I'll get back to you.
ashishjaddu@gmail.com
Connect with me