A next‑generation, Azure‑powered, multi‑agent orchestration system designed to automate IT, HR, Finance, and operational workflows using structured agents, deterministic runbooks, and intelligent routing.
Built collaboratively by the TriResolve AI Team for the 2025 AI Engineering Hackathon.
- 🤖 Multi-Agent Intelligence across IT, HR, Finance, Ops, Security, and Architecture domains
- 🧭 TriNexa Orchestrator for multi-agent coordination and final answer synthesis
- 🧠 Classifier Agent for ticket domain + intent prediction
- 📘 Deterministic YAML Runbooks for consistent, auditable workflows
- ⚡ FastAPI Backend with Azure OpenAI integration
- 🔍 Observability & Logging for agent calls and orchestration traces
- 📊 Streamlit Frontend for visual demo + interactive UX
- 🐳 Docker-ready for reproducible deployment
The TriResolve AI platform is built using a layered architecture:
flowchart TD
UI([Streamlit Frontend]) --> API(FastAPI Backend)
API --> Classifier>Classifier Agent]
API --> Orchestrator>TriNexa Orchestrator]
Orchestrator --> IT[IT Agent]
Orchestrator --> HR[HR Agent]
Orchestrator --> FIN[Finance Agent]
Orchestrator --> ARCH[Architect Agent]
Orchestrator --> SEC[Security Agent]
Orchestrator --> OPS[Ops Agent]
IT --> RB((Runbooks))
HR --> RB
FIN --> RB
OPS --> RB
Orchestrator --> Final[(Final Response)]
Final --> UI
Handles:
- Ticket intake and validation
- Communication with Azure OpenAI agents
- TriNexa Orchestrator routing logic
- Response standardization
Each agent is defined with:
- Instructions (persona + rules)
- Input + Output schemas
- Optional tools (KB search, runbooks, etc.)
- IT Agent – device + access troubleshooting
- HR Agent – onboarding + employee policies
- Finance Agent – invoices, vendor, reimbursements
- Security Agent – risk review + compliance rules
- Ops Agent – SRE-style incident triage
- Architect Agent – system design + solution planning
- Classifier – domain + intent prediction
- TriNexa Orchestrator – multi-agent routing + final answer synthesis
Stored under /agents/<domain>/runbooks/.
YAML-based deterministic actions that agents reference to:
- Structure multi-step processes
- Maintain auditability
- Prevent inconsistent behavior
Intuitive interface for demo and judging:
- Map page for agent architecture visualization
- Assistant page for real-time interactions
- Color‑coded departmental themes
- Runs locally or via Streamlit Cloud
TriResolve AI is built with:
- Python 3.11+
- FastAPI
- Streamlit
- Azure OpenAI (multi-agent deployments)
- Azure AI Foundry (agent instructions + schemas)
- Azure Cognitive Search (optional KB integration)
- Docker
- GitHub Actions
.env+ GitHub Secrets
- Synthetic ticket datasets (IT/HR/Finance)
- Foundry metadata for agents
- YAML runbooks
triresolve-service-desk/
├── agents/
│ ├── it/
│ ├── hr/
│ ├── finance/
│ ├── security/
│ ├── ops/
│ ├── architect/
│ └── orchestrator/
│
├── backend/
│ ├── api/
│ ├── services/
│ ├── schemas.py
│ ├── azure_client.py
│ └── orchestrator.py
│
├── streamlit/
│ ├── streamlit_app.py
│ └── pages/
│
├── docs/
├── runbooks/
├── .env.example
├── Dockerfile
└── requirements.txt
uvicorn backend.api.main:app --reloadstreamlit run streamlit/streamlit_app.pydocker build -t triresolve-ai .
docker run -p 8000:8000 triresolve-aiAPI docs available at:
http://localhost:8000/docs
-
Purpose: Run the backend locally without Azure OpenAI credentials. When
TRIRESOLVE_DEV_MODEis enabled the backend will return canned, deterministic responses useful for UI development and testing. -
Environment variables:
TRIRESOLVE_DEV_MODE— set to1,true, oryesto enable dev mode.TRIRESOLVE_CANNED_CLASSIFICATION— optional JSON string returned for classifier-style prompts (example:{"department":"IT","confidence":0.95,"rationale":"..."}).TRIRESOLVE_CANNED_REPLY— optional short text reply used for general orchestrator/domain responses.
-
Run server in dev mode:
export TRIRESOLVE_DEV_MODE=true # optional overrides export TRIRESOLVE_CANNED_CLASSIFICATION='{"department":"IT","confidence":0.95,"rationale":"Dev mode"}' export TRIRESOLVE_CANNED_REPLY='DEV: we will handle this shortly.' nohup python -m uvicorn backend.api.main:app --host 0.0.0.0 --port 8000 > uvicorn.log 2>&1 & echo $! > uvicorn.pid tail -f uvicorn.log
-
Run tests (including dev-mode tests):
pytest -q
This makes it easy to iterate on frontend and orchestration code without requiring access to Azure OpenAI during initial development.
TriResolve AI was developed using strong Responsible AI principles to ensure safety, security, compliance, and ethical automation.
- Zero real user data used; 100% synthetic datasets
- Secrets isolated in
.envand GitHub Secrets - Azure RBAC enforced across OpenAI and Foundry resources
- Logs avoid collecting PII and redact sensitive content
- Domain-specific allow/deny lists embedded in each agent
- Orchestrator enforces safe multi-agent routing
- Security Agent evaluates identity, access, and compliance risks
- Finance, HR, and IT agents strictly follow policy-aligned boundaries
- Orchestration logs show exactly which agents were called
- Deterministic YAML runbooks create consistent, auditable outcomes
- Each agent returns structured fields for full traceability
- Designed to augment—not replace—human decision-making
- Automated actions include built-in escalation pathways
- Security, Ops, and HR agents elevate risk-sensitive requests to humans
- No private or organizational datasets used
- Synthetic tickets generated for IT, HR, and Finance
- Dataset bias manually checked and minimized
TriResolve AI demonstrates:
- Fully working multi-agent orchestration
- JSON-standardized outputs
- Structured runbook-driven automation
- Full audit + reasoning trace through TriNexa
- A polished end-to-end UX
- M1 – Foundations
- M2 – Agents + Backend Routing
- M3 – Classifier Integration
- M4 – Demo UX + Streamlit
- M5 – End-to-End System Demo
- M6 – Final Submission
A globally distributed engineering team:
Lead Architect & AI Systems Designer
Backend Developer & Data Engineer
Machine Learning Engineer
UX Contributor & Logic Reviewer
TriResolve AI was created through collaborative engineering, multi‑agent experimentation, and cross‑domain product design by the TriResolve team.
The system is designed for extensibility, auditability, and real-world enterprise service desk automation.