Aman Saraswatiitg
Agentic AI Engineer from IIT Guwahati — multi-agent systems
New to Contra
Aman is building their profile!
Built a 4-agent e-commerce operations swarm where every
agent has one job — and they do it in real-time, together.
Most e-commerce teams manage inventory, pricing, fraud,
and customer ops in silos. This system connects all 4
into one intelligent operation:
→ Inventory Agent
Stock monitoring, demand forecasting, reorder
automation — no more stockouts
→ Pricing Agent
Real-time competitor tracking, margin intelligence,
dynamic price optimization across 8 SKUs
→ Fraud Detection Agent
Anomaly detection, risk scoring, velocity checks,
IP blacklist, COD pattern analysis — stops fraud
before it happens
→ Customer Operations Agent
Ticket resolution, order tracking, sentiment
analysis — faster responses at scale
All 4 orchestrated by a Supervisor Agent that
coordinates, validates, and ensures business goals
are met — not just tasks completed.
What makes it production-ready:
✦ EventBus architecture for real-time agent communication
✦ WebSocket-powered live dashboard (glassmorphism UI)
✦ 40 automated tests covering all agent behaviors
✦ ERP · WMS · CRM · Orders · Payments data integration
✦ End-to-end audit trail from ingestion to action
Stack: Python · FastAPI · WebSocket · Chart.js ·
CrewAI · OpenAI API
1
11
Built an RBI-compliant RAG system with enterprise-grade
security — because "it answered correctly" isn't enough
in regulated industries.
Most RAG systems stop at retrieval. This one has 3
security layers before a single response reaches the user:
→ Layer 1: Adversarial prompt protection
Lexical regex filtering catching jailbreak attempts
→ Layer 2: Semantic cosine-similarity detection
Catches sophisticated prompt injection that bypasses
keyword filters
→ Layer 3: Canary token leak validation
Detects if sensitive knowledge base content is
being extracted
What else is built in:
✦ FAISS vector search over RBI & fintech policy docs
✦ PII masking — Aadhaar & PAN auto-detected & masked
✦ SHA-256 audit trail on every single query
✦ Explainable AI — every response shows sources +
confidence + latency
✦ Human-in-the-loop review for high-risk queries
The result: A RAG system you can actually deploy in
a bank — not just a demo.
Stack: Python · FAISS · Sentence-Transformers ·
ChromaDB · Streamlit · OpenAI API
1
28
Built a production-grade 5-agent AI orchestration system
deployed on Kubernetes — not a prototype.
The system runs 5 specialized agents in parallel:
→ Planner Agent: Breaks down complex tasks
→ Research Agent: Retrieval and context gathering
→ Executor Agent: Core task execution
→ Reviewer Agent: Quality validation
→ Guardrail Agent: Compliance + safety enforcement
What makes it enterprise-grade:
✦ Kubernetes HPA auto-scaling (2–10 replicas)
✦ EU AI Act + GDPR compliance guardrails
✦ Prometheus/Grafana full observability stack
✦ Human-in-the-loop approval for high-risk decisions
✦ SHA-256 audit trails on every agent action
✦ PII detection and masking via Guardrails AI
Stack: Python · CrewAI · FastAPI · Docker · Kubernetes ·
Prometheus · Grafana · Guardrails AI · OpenAI API
0
12