Hitarth Bharad

Hitarth Bharad

Software Engineer — specializing in ML & AI Systems

Latest: Releasing early version of Aegion
Software EngineerLLM SystemsVector SearchProduction ML
View Aegion Case StudyDownload Resume
aegion-cli
$ aegion run \
  "Fetch reviews from CRM, score sentiment, email report"

[discover] Found 4 tools in registry
[plan]     Building workflow DAG...

// Workflow chain
{
  "steps": [
    "CRM API" → "Sentiment Scorer" → "Report Writer" → "Email API"
  ],
  "status": "complete",
  "time": "1.2min",
  "confidence": 0.94
}

Featured Work

Aegion

AI-Powered Workflow Orchestration — Let LLMs Build and Execute Complex Business Workflows

Tools Registered

100+

across users

Workflow Success Rate

94%

auto-generated chains

Avg Workflow Time

3min

vs 45min manual

Tool Chain Accuracy

96%

correct tool selection

Overview

Aegion transforms natural language requests into executable workflows by intelligently orchestrating user-registered tools. Users register their own APIs, Python functions, and LLM agents as tools in their library. When they ask a question like “Fetch customer reviews from our CRM, analyze sentiment, and email a report,” Aegion’s AI automatically identifies the required tools, chains them in the correct sequence, and executes the complete workflow—no manual pipeline building required.

Example Workflow

User Input:
"Fetch all customer reviews from CRM in past 7 days,
 compute confidence scores, email me a detailed report"

Aegion Process:
1. Identifies tools: CRM API → Python Scoring → LLM Report Writer → Email API
2. Builds workflow chain with proper data flow
3. Executes pipeline and delivers results

Result: 45-minute manual task → 1-minute automated workflow

System Architecture

How It Works

01

Register Tools

  • •Add existing APIs, Python functions, or configure LLM agents
  • •Provide tool name, description, input/output schema, auth
  • •Aegion indexes each tool for semantic search
02

Natural Language Request

  • •Describe workflow in plain English: "Fetch reviews → Analyze → Report → Email"
  • •No manual pipeline configuration required
  • •AI identifies required tools automatically
03

Automated Execution

  • •Aegion builds the workflow chain automatically
  • •Executes tools in sequence with proper data passing
  • •View results and full execution trace

Technical Deep Dive

Intelligent Tool Discovery & Selection

Challenge: Given a natural language request, how do you automatically identify which tools (from potentially hundreds in a user’s library) are needed and in what order?

Solution: Vector embeddings of all registered tools (using tool descriptions, parameters, outputs). Semantic search to find relevant tools based on user intent. LLM-powered reasoning to understand tool dependencies and data flow requirements. Confidence scoring for tool selection (reject if confidence < 85%).

92%

tool selection accuracy

2.3s

discovery in 100+ tool libraries

5-8

multi-step tool sequencing

# Tool matching with semantic search
query_embedding = embed_user_intent(user_request)
candidate_tools = vector_db.similarity_search(
    query_embedding, 
    top_k=10,
    threshold=0.75
)

# LLM validates and orders tools
workflow_chain = llm.plan_workflow(
    user_request=user_request,
    available_tools=candidate_tools,
    validate_dependencies=True
)

Workflow DAG Generation & Validation

Challenge: Automatically building a valid execution graph where tools pass data correctly between steps, handling different output/input formats.

Solution: Parse tool schemas (input params, output types) from registrations. LLM generates directed acyclic graph (DAG) with data mappings. Static validation checks for type compatibility. Automatic data transformation injection where needed (JSON → CSV, etc.).

94%

workflows execute without errors

Branching

conditional tool execution

4.2

avg tools per workflow chain

Example DAG

1Fetch CRM DataAPIreturns: JSON array
2Score ReviewsPythonreturns: DataFrame
3Generate ReportLLM Agentreturns: Markdown
4Send EmailAPIreturns: success

Tool Type Flexibility & Execution

Challenge: Supporting three fundamentally different tool types (APIs, Python code, LLM agents) in a unified execution framework.

API Tools

REST/GraphQL wrapper with authentication, rate limiting, retry logic

Python Tools

Sandboxed execution environment with dependency management and timeout controls

LLM Agents

Completion agents with role/context injection, streaming support, cost tracking

Execution Service Features

  • •Async/parallel execution where dependencies allow
  • •State management for long-running workflows
  • •Real-time execution monitoring and logging
  • •Automatic error recovery with fallback tools

<5ms

overhead per tool execution

99.2%

workflow completion rate

Graceful

degradation on tool failures

Technology Stack

TechnologyRationale
Vector DatabasePinecone — Semantic tool search with <100ms query time across 10K+ tool embeddings
Workflow PlanningGPT-4 + Claude — Multi-model approach for higher accuracy in tool selection and DAG generation
Execution EngineCustom orchestrator on AWS ECS — Async task queue with state management and retry logic
Tool SandboxDocker + gVisor — Secure isolated execution for user Python code with resource limits
Tool RegistryPostgreSQL + Redis — Fast tool lookup with caching

Business Impact

Time Savings

45 min → 3 min

average for multi-step workflows

Adoption

100+

tools registered across users

Reliability

94%

workflow success rate on first generation

Flexibility

3 types

API, Python, LLM agents

Developer Experience

Low-code

workflow automation for business users

Try Aegion

Production Systems

Other Shipped Work

Auto PO Processor

Intelligent document processing platform extracting structured data from complex PDFs with OCR and NLP.

Accuracy95%Multi-format PDFs
  • •Achieved 95% extraction accuracy across multi-format document types using Tesseract OCR and custom text parsing algorithms
  • •Built secure processing pipelines with PII data encryption and automated quality assurance workflows
  • •Implemented comprehensive testing suite with PyTest ensuring data integrity and validation across edge cases

Python, FastAPI, Next.js, TypeScript, Tesseract OCR, ShadCN UI, PyTest

View Details →

Tucson Crime Pattern Analysis Dashboard

End-to-end analytics platform processing real-time crime data with predictive modeling and geospatial visualization.

Uptime99.9%Real-time ETL
  • •Built automated ETL pipeline with 99.9% uptime fetching and processing crime data from Police Department REST APIs
  • •Developed interactive geospatial dashboard using R Shiny and Leaflet.js for real-time crime hotspot visualization with multi-dimensional filtering
  • •Applied time-series forecasting (ARIMA) and spatial clustering models enabling data-driven resource allocation for law enforcement

R (Shiny, Quarto), PostgreSQL, Leaflet.js, REST APIs, ARIMA, Spatial Analysis

View Details →

Anomaly Detection in Time-Series Patterns

Production anomaly detection pipeline processing 1M+ data points daily from financial and sensor sources.

Detection Precision92%-35% response time
  • •Developed automated data pipeline with quality checks processing 1M+ daily data points from financial and sensor streams
  • •Achieved 92% precision in anomaly detection using ensemble of ARIMA, EMA, and Prophet models with minimal false positives
  • •Built interactive visualization dashboards reducing incident response time by 35% through actionable insights

Python, Pandas, NumPy, Scikit-learn, Prophet, ARIMA, PostgreSQL, Matplotlib, Seaborn

View Details →

VSI Lab Portal

Centralized research management ecosystem for the Vision Systems and Intelligence Lab, streamlining academic profiles and publication tracking.

Global ReachHigh Impactfrom Zero Presence
  • •Architected a research repository using Next.js 14 and Vercel, achieving sub-second page loads for large publication datasets
  • •Engineered a responsive, accessible UI/UX framework using Tailwind CSS and Framer Motion for seamless navigation across scholarly content
  • •Implemented a dynamic data-fetching layer to synchronize lab member profiles, research projects, and peer-reviewed publications

Next.js, TypeScript, React, Tailwind CSS, Framer Motion, Vercel

View Details →

Research

Publications

Research informing production work

FastSpectralNet: Efficient Hyperspectral Image Classification using Context Vector Attention

H. Bharad, J. Biradar, E. Lee • WACV • 2025

Under Review

This paper explores novel Vision Transformer architecture for Hyperspectral Image Classification which provides low latency and higher accuracy for inference