Experimental AI Engineering

AI Proof of Concept
Projects

A collection of hands-on AI systems built to explore LLM workflows, RAG pipelines, agentic architectures, and serverless AI execution on AWS. Each PoC targets a real-world problem and demonstrates a distinct AI engineering pattern.

View on GitHub Back to Portfolio
0
PoC Projects
0
AI Patterns
0
Cloud Services
// poc projects

AI Systems Built

Four independent systems, each exploring a different AI engineering discipline.

🧠
AI Meeting Notes Summarizer
A production-ready FastAPI + LangChain service that ingests meeting transcripts and automatically generates structured notes — summaries, key decisions, and action items with owners and deadlines.
Summarization Vector Store
Key Capabilities
  • Accepts .txt transcript uploads via REST API endpoint
  • Splits and embeds transcript chunks into a Chroma vector store using local HuggingFace embeddings (all-MiniLM-L6-v2)
  • Structured output via Pydantic parser — summary, decisions, and action items with owner & deadline fields
  • LLM-powered via google/gemma-3-12b-it:free through OpenRouter with automatic retry & backoff
  • Fully containerised with Docker Compose (API + Chroma services)
Tech Stack
Python FastAPI LangChain Chroma DB HuggingFace OpenRouter LLM Docker Pydantic
Real-World Use Cases
Corporate meeting automation CRM action item extraction Legal transcript processing HR interview notes
Architecture Flow
POST /upload-transcript → TextLoader → RecursiveCharacterTextSplitter → Chroma Vector Store → PromptTemplate → ChatOpenAI (OpenRouter) → PydanticOutputParser → MeetingOutput JSON
silypointer/ai_meeting_notes_summerizer
LangChain Pipeline Structured Output Dockerised
🔍
Enterprise Policy Q&A — RAG System
A locally-deployed Retrieval-Augmented Generation pipeline for querying enterprise policy PDFs. Ask natural language questions and receive grounded answers with source citations and LLM-as-judge evaluation metrics.
RAG Hybrid Retrieval
Key Capabilities
  • PDF ingestion via PyMuPDF — parsed, chunked (512 chars / 64 overlap), and embedded locally
  • Hybrid retrieval: dense vector search (Qdrant + BAAI/bge-small-en-v1.5) fused with BM25 keyword search via Reciprocal Rank Fusion (RRF)
  • Cross-encoder reranking with ms-marco-MiniLM-L-6-v2 for precision-ranked results
  • Optional LLM query rewriting to expand ambiguous questions before retrieval
  • LLM-as-judge evaluation: faithfulness, context precision, context recall, and answer relevance (scored 0–1)
Tech Stack
Python FastAPI Qdrant BM25 Cross-Encoder PyMuPDF HuggingFace OpenRouter LLM Docker
Real-World Use Cases
HR policy chatbot Legal document Q&A Compliance assistant Internal knowledge base
Architecture Flow
PDF Upload → PyMuPDF Parser → Text Chunker → Qdrant (dense) + BM25 (keyword) → RRF Fusion → Cross-Encoder Reranker → LLM Answer Generator → Answer + Citations
silypointer/enterprise_policy_qa_rag
Hybrid RAG Reranker LLM Evaluation
⚡
AI Lambda Functions
A collection of serverless AWS Lambda functions that expose AI capabilities — text generation, AI image creation, a conversational chatbot, and a RAG prototype — all powered by Amazon Bedrock foundation models.
Serverless AWS Bedrock
Functions Included
  • textGenerationFunction — Invokes Bedrock Titan Text Express to generate creative or business content from a prompt. Configurable temperature, topP, and token count.
  • moviePosterDesignFunction — Generates AI movie poster images via Bedrock Titan Image Generator, uploads to S3, and returns a pre-signed URL (30-min expiry).
  • chatbot — Streamlit + LangChain conversational chatbot backed by AWS Bedrock (Amazon Nova Pro / DeepSeek). Maintains conversation context via ConversationSummaryBufferMemory.
  • rag — Lightweight RAG prototype with document loading, text splitting, and retrieval-augmented Q&A using Bedrock models.
Tech Stack
Python AWS Lambda AWS Bedrock Titan Text Express Titan Image Generator Amazon S3 LangChain Streamlit boto3
Real-World Use Cases
Serverless content generation AI image creation APIs Embedded chatbots Event-driven AI pipelines
Architecture Flow
HTTP Event / Trigger → AWS Lambda Handler → Input Validation → Bedrock InvokeModel (boto3) → Response Parsing → S3 Upload (image) / JSON Response (text)
silypointer/ai_lambda_functions
Text Generation Image Generation Chatbot RAG Prototype
🤖
Node.js AI Agent — ToDo Manager
A terminal-based agentic AI application that accepts natural language instructions, maps intent to safe pre-defined database operations, and manages a PostgreSQL-backed todo list — without ever executing LLM-generated SQL.
AI Agent Tool Calling
Key Capabilities
  • Natural language CRUD — create, list, update, delete, and search todos via conversational input
  • Strict tool registry: LLM can only invoke one of 6 pre-defined tools (create_todo, list_todos, update_todo, delete_todo, mark_completed, search_todos)
  • All database operations use parameterised SQL — no raw LLM-generated queries ever reach the database
  • Runtime conversation memory maintains context across the session
  • OpenRouter integration with retry logic, structured logging, and configurable model selection
Tech Stack
Node.js OpenRouter LLM PostgreSQL Tool Calling Parameterised SQL UUID Keys Conversation Memory
Real-World Use Cases
AI-powered task managers Natural language DB interfaces Safe agentic CRUD systems Internal ops automation
Architecture Flow
User Natural Language Input → Agent Service → OpenRouter LLM (tool_call) → Tool Registry Validation → Repository Function → Parameterised SQL → PostgreSQL → Conversational Response
silypointer/nodejs_ai_agent
Agentic Architecture Safe Tool Calling Conversation Memory
// engineering depth

AI Engineering Areas Explored

Patterns and disciplines applied across these PoC projects.

📚
Retrieval-Augmented Generation (RAG)
Hybrid dense + keyword retrieval, RRF fusion, cross-encoder reranking, and grounded answer generation with source citations.
🎯
LLM Prompt Engineering
Structured output prompting with Pydantic parsers, query rewriting, system prompt design, and few-shot instruction patterns.
🤖
Agent-Based Workflows
Tool-calling agents with strict registry validation, conversation memory, safe parameterised database execution, and retry logic.
⚡
Serverless AI Execution
AWS Lambda functions invoking Bedrock foundation models for text and image generation, with S3 storage and pre-signed URL delivery.
📝
Automated Summarization Pipelines
End-to-end transcript ingestion, vector embedding, LangChain orchestration, and structured JSON output extraction for meeting notes.
🧪
LLM Evaluation (LLM-as-Judge)
Automated quality scoring using a secondary LLM to measure faithfulness, context precision, context recall, and answer relevance.
// get in touch

Interested in Collaborating?

Open to consulting roles, freelance AI projects, and interesting engineering challenges. Drop me a message — I usually respond within 24 hours.

📧
Email
silypointer@gmail.com
🐙
GitHub
github.com/silypointer
🗂️
Portfolio
View Full Portfolio