🏢 Zillion Technologies

2024 – Jan 2026 | Hybrid

---

Software Engineering Intern — Summer 2024

Project: Intelligent File Search & Vector Database Integration

When I joined Zillion Technologies as an intern, most interns were assigned exploratory or future-looking projects. Thanks to my full-stack experience, I was entrusted with a live, client-facing deliverable — a system that transformed SharePoint file repositories into an intelligent, searchable vector database.

• Optimized SharePoint API Integration
  • Took over existing implementation and significantly improved performance and structure
  • Enhanced secure auth (secret ID, secret key, user token), recursive file navigation, large-volume handling
• Vector Embedding for Semantic Search
  • Designed pipeline for document content (OCR for image-based PDFs) → vector database
  • Used all-MiniLM-L6-v2 transformer for embeddings
• Chatbot Interface
  • Integrated OpenAI ChatGPT for conversational queries
  • Results included content matches, file locations, and metadata for precise navigation
• Delta Updating
  • Implemented selective re-indexing of only modified files/folders
  • Reduced processing time and improved scalability

---

Software AI Engineer Intern — Summer 2025

Project: Zecured — Admin Panel

Worked on an entirely new feature for Zecured: the Admin Panel. The panel consisted of three main flows.

• Onboarding & Offboarding

  • Given credentials, onboard and offboard users from IAM systems
  • Streamlined access management for enterprise environments

• Governance — Role Clustering

  • For companies with hundreds or thousands of roles (each with entitlements, role names, job descriptions), the system matches and buckets similar roles based on entitlements (highest), descriptions, and other attributes
  • Regex matching for entitlements and descriptions to form and compare buckets
  • Users can approve a bucket, reject it, assign a new role name, or update descriptions
  • Reduced role sprawl and simplified access governance

• AI-Suggested Access Level

  • AI runs through each role’s description and suggests an access level — which can differ from the role’s current access level
  • From a bucket of many options (e.g., 10), you choose a subset (e.g., 3): for each you pick either the suggested access level or keep the original (as long as they match where needed)
  • On confirm: one merged role is created (one new role bucket); the remaining roles form new buckets
  • Lets you consolidate roles with consistent, AI-informed access levels while keeping flexibility

• Approval Flow

  • Two-step approval: PO approval + Manager approval
  • Both approvals = green light; any rejection = red light
  • Clear visibility into approval status across the workflow

---

What I learned (Zecured / IAM)

• IAM in the enterprise — How companies structure identity and access: onboarding/offboarding, entitlements, role hierarchies, and how governance panels sit on top of existing IAM systems
• Regex for entitlements and descriptions — Matching and bucketing roles using patterns on entitlement strings and job descriptions so similar roles cluster correctly for merge or review

---

AI Engineer

Project 1: Agentic AI Voice Agent — Sysco

Built an agentic AI system implemented in LangGraph and Google ADK (experimented with both). The project was under Sysco (the food company), with order and product data in Salesforce and authentication by phone number.

• Connected Salesforce to the agent — Pulled orders, people, contacts, and items so the voice agent could answer questions about real customer data. The agent queried Salesforce in real time to respond to natural language requests.

The system was a voice-enabled answering machine: users could call in and check on their orders using natural language. Examples: "Tell me about my last order", "What was the order we just talked about?", "My order with a burger". Connected to Talkdesk via websockets for real-time voice.

What I learned (Agentic AI)

• Google’s agent convention — A top-level agent orchestrates a sequential agent and tools; tools can invoke further agents, almost like nodes in a graph. This funneling gives the LLM deterministic paths and hard data (e.g., from Salesforce) instead of free-form generation, which reduces hallucinations and keeps answers grounded.
• LangGraph / Google ADK — Structuring multi-step, tool-calling flows so the model follows a controlled path rather than “making it up.”
• Operational — Google Cloud Console, enabling APIs, Vertex AI for deployment; Talkdesk (flows, call routing, IVR, webhooks); Salesforce (orders, people, contacts, items, API).

---

Project 2: Local LLM on RTX 3060 — Phi-3 Mini

Ran Phi-3 Mini locally on an RTX 3060. Quantized in 8-bit and 4-bit to fit in 3GB and 5.6GB VRAM respectively, and configured all libraries (CUDA, transformers, bitsandbytes, etc.) to use the GPU.

Implemented this in parallel in Zecured to generate entitlement descriptions. A parallel tab compared the local LLM output to GPT — performance was competitive. Benefits: no API calls to OpenAI, ensuring data privacy and reducing costs.

What I learned (Local LLMs)

• Temperature & sampling — Temperature controls randomness; do_sample (or equivalent) makes the model more creative vs deterministic. Tuning these changes how “strict” or varied the outputs are.
• 4-bit vs 8-bit quantization — Ran 4-bit (e.g., ~3GB VRAM) and 8-bit (e.g., ~5.6GB VRAM) on my own hardware (RTX 3060) to speed up inference and fit larger contexts; traded a bit of quality for latency and memory.
• Stack — Local LLM deployment, CUDA/GPU setup, Hugging Face transformers, bitsandbytes; getting Phi-3 Mini running and integrated into Zecured.

---

🔧 Technologies Used

Summer 2024 — SharePoint Vector Search:

• Languages & Frameworks: Python (FastAPI)
• APIs: Optimized SharePoint API
• LLMs: OpenAI ChatGPT
• Embeddings: Hugging Face all-MiniLM-L6-v2
• Vector Database: ChromaDB
• OCR: Tesseract
• Other: File system crawling, chatbot integration, metadata parsing

Summer 2025 — Zecured Admin Panel:

• IAM onboarding/offboarding, role governance, approval workflows
• Regex matching for entitlements/descriptions; AI-suggested access level, merge/new buckets

AI Engineer (2025 – Jan 2026):

• Agentic AI: LangGraph, Google ADK, Vertex AI, Google Cloud Console
• Voice: Talkdesk (websockets, flows, call routing, IVR), Salesforce (orders, contacts, items, API)
• Local LLM: Phi-3 Mini, 4-bit/8-bit quantization, RTX 3060 / CUDA
• Stack: Hugging Face transformers, bitsandbytes, GPU-optimized inference
• Zecured: Entitlement description generation, parallel local vs. cloud LLM comparison

---

💡 Why It Mattered

This wasn’t a sandbox — it was a production-grade, client-facing system. Hands-on experience in:

• Improving real-world API integrations
• Scalable document ingestion and vectorization
• NLP-powered semantic search
• Delta indexing for efficiency
• User-focused AI applications for enterprise

Outcome: Delivered a working prototype enabling real-time, context-aware document retrieval from complex SharePoint file systems.

---