Founding AI Engineer
Taj Khunkhun
Transforming experimental prototypes into production-grade agentic AI systems. 8+ years in backend systems, 3+ years specializing in Multi-Agent Orchestration and Autonomous Reasoners.
LangGraph · CrewAI · AutoGen · MCP · RAG
What I Work With
Technical Skills
AI & Agents
ML & Deep Learning
Languages & Frameworks
Data & Infrastructure
Cloud & DevOps
Where I've Worked
Experience
Lextar AI Legal Solutions, Inc.
Founding AI Engineer
Remote
- ▸Architected a full-stack enterprise legal AI platform for Canadian and US immigration law using FastAPI, Next.js, PostgreSQL, SQLAlchemy, and ChromaDB for hybrid RAG.
- ▸Engineered an 11-step lawyer-faithful reasoning pipeline integrating DeepSeek LLM, hierarchical issue framing, conflict resolution, and strict grounding verification.
- ▸Built a hybrid GraphRAG retrieval system combining ChromaDB vector search with an in-memory knowledge graph (1,300+ legal nodes, 445 relationships) across 6,200+ ingested chunks.
- ▸Implemented multi-tenant SaaS with Clerk authentication, role-based access control, Stripe checkout, PDF invoices, and automated SMTP email receipts.
- ▸Designed superadmin dashboard with platform-wide analytics, user management, per-organization usage tracking, and real-time reasoning unit consumption monitoring.
- ▸Built real-time streaming legal analysis workspace with SSE, live reasoning trace visualization, and role-specific output formatting.
- ▸Implemented strict legal output governance including grounding verification, cross-reference resolution, and confidence-weighted scoring.
Product Demo
Hewlett Packard Enterprise
Senior AI Engineer | Agentic AI Engineer
Spring, Texas (Remote)
- ▸Designed multi-agent Plan-and-Execute architecture using LangGraph, reducing infinite loop failures by 35%.
- ▸Architected MCP server layer standardizing tool integration, reducing custom tool-binding code by 60%.
- ▸Reduced inference costs by 45% via Router Agent dynamically triaging between Llama 2 7B and GPT-4.
- ▸Implemented HITL checkpoint system for financial workflows with 0.85 confidence threshold.
- ▸Built observability dashboards using LangSmith and Arize Phoenix, resolving 10s+ latency bottlenecks.
- ▸Architected LLM evaluation framework using DeepEval across 5,000+ test cases.
- ▸Reduced costs by 52% and p95 latency by 300ms through semantic caching with Redis.
- ▸Developed Shadow Deployment pipeline for prompt engineering A/B tests on live traffic with 0% user impact.
- ▸Designed multi-agent collaboration framework with Planner, Executor, and Critic agents via structured message-passing.
- ▸Eliminated catastrophic forgetting in Llama 2 by mixing 15% pre-training replay data.
- ▸Reduced training VRAM by 65% using QLoRA 4-bit quantization for 70B parameter models.
- ▸Improved inference throughput 4x via multi-LoRA serving (vLLM/LoRAX) for 10+ adapters.
- ▸Architected multi-region Kafka with sub-second cross-region replication.
- ▸Engineered tiered memory system preserving intent across 20+ agent handoffs.
- ▸Led migration from monolithic Django to Microservices with Docker and Kubernetes.
- ▸Built high-concurrency FastAPI architecture handling 10k+ concurrent WebSocket connections.
- ▸Built MCP-compliant tool registry enabling dynamic tool discovery at runtime.
- ▸Architected Self-Correction loop for SQL agent, reducing syntax errors by 50%.
- ▸Built NER pipeline with Keras Bi-LSTM achieving 25% F1-score improvement.
- ▸Resolved critical GIL bottlenecks by refactoring to Python Multiprocessing and Celery/Redis worker cluster.
- ▸Engineered custom spaCy and Transformer models integrated with Fabric Lakehouse using lazy-loading across Spark clusters.
Adobe
Data & Machine Learning Engineer
San Jose, California
- ▸Architected ETL pipeline processing 5TB+ multi-modal data using Spark.
- ▸Eliminated vector-relational desync via CDC (Debezium + Kafka) with 99.9% consistency.
- ▸Engineered Blue-Green re-indexing for zero-downtime migrations across 50M+ vectors.
- ▸Optimized semantic search by 40% via hierarchical document indexing.
- ▸Reduced vector storage costs by $8k/month through tiered data strategy.
- ▸Deployed Semantic Data Guard monitoring data drift with 15% deviation alerting.
- ▸Standardized AI Data Contracts across four teams enforcing GDPR/CCPA compliance.
- ▸Reduced inference latency by 65% via model distillation on NVIDIA A100 GPUs.
- ▸Built synthetic data engine using SDV and GPT-3.5, improving minority tasks by 18%.
- ▸Engineered Fail-Soft orchestration saving $15k/month in compute costs.
- ▸Solved multi-modal cold start problem with tiered embedding cache (Redis + Elasticsearch), reducing first-response time from 2.4s to 400ms.
- ▸Built programmatic A/B testing framework promoting models based on Ground Truth alignment with statistical significance tests.
Enterprise Work
Projects
Full-stack legal AI SaaS platform with hybrid GraphRAG retrieval (1,300+ node knowledge graph + vector search), 11-step structured legal reasoning pipeline, multi-tenant RBAC, real-time streaming analysis workspace, and governance-grade output verification for Canadian and US immigration law.
Multi-agent Plan-and-Execute architectures, Router Agents, MCP server integrations, HITL checkpoints, semantic caching, shadow deployment pipelines, and LLM evaluation frameworks for HPE's enterprise AI platform.
Fine-tuned LLMs with QLoRA on GPU clusters, multi-region Kafka replication, NER pipelines, multi-agent memory systems, and multi-LoRA inference serving across Kubernetes-based ML platform.
Migrated monolithic services to microservices, high-concurrency event-driven APIs, MCP-compliant tool registries, SQL-generating agents, and GIL/integration bottleneck resolution.
High-throughput ETL/search pipelines, CDC-based vector sync, embedding drift management, synthetic data generation, inference optimization, and A/B testing frameworks.
Open Source
Side Projects
Personal projects exploring agentic AI patterns, multi-agent architectures, and intelligent automation.
Agentic AI Chat Analyzer
AI-powered platform for analyzing agent chat transcripts. Performs exploratory data analysis, LLM-based summarization, and sentiment classification through an interactive Streamlit frontend.
- ▸Modular data pipeline (ingestion, cleaning, transformation)
- ▸EDA with word clouds and sentiment visualizations
- ▸Model caching for offline operation
AI Recruiter
Intelligent recruitment platform that automates candidate discovery by scanning GitHub profiles and Google Scholar to identify qualified AI/ML professionals with relevance scoring.
- ▸Multi-source profile analysis (GitHub + Google Scholar)
- ▸Relevance-based intelligence scoring for AI/ML skills
- ▸Geographic filtering and co-author extraction
AI Email Agent (Supervisor Mode)
Email automation system using supervisor-pattern multi-agent architecture that categorizes emails, generates RAG-powered responses, proofreads with AI, and sends replies via Gmail.
- ▸Supervisor pattern for dynamic agent coordination
- ▸RAG-powered response generation from knowledge base
- ▸AI proofreading layer before sending
Simple Chatbot
Lightweight, rule-based chatbot with Gradio UI that answers questions about healthcare automation agents using fuzzy string matching -- no external LLM calls required.
- ▸Weighted scoring: string similarity (60%) + keyword matching (40%)
- ▸Confidence threshold for answer selection
- ▸Graceful fallback responses listing available topics
Academic Background
Education
Master of Computer Science
Santa Clara University
Bachelor of Computer Science
Santa Clara University
Let's Connect
Get in Touch
I'm always open to discussing new opportunities in Agentic AI, Multi-Agent Systems, and production ML engineering.