Dileepa Dilshan DileepWick

DILEEPA DILSHAN

AI ENGINEER | SOFTWARE ENGINEER • HASHBAZE • COLOMBO, SRI LANKA

ENGINEERING PRODUCTION AI SOFTWARE WITH MODERN SOFTWARE ENGINEERING PRINCIPLES

// 01 • PROFILE & CORE FOCUS

I am Dileepa Dilshan, an AI Engineer and Software Engineer based in Colombo, Sri Lanka, working at Hashbaze. I design and build production-grade software that bridges robust software engineering principles with state-of-the-art artificial intelligence.

My engineering practice emphasizes clean architecture, type safety, deterministic API design, and performance optimization across TypeScript and Python. On the AI side, I build vector search pipelines, LLM security guardrails against prompt injection, Byte-Pair Encoding (BPE) token labs, and Model Context Protocol (MCP) integrations.

TECHNICAL FOCUS & COMPETENCIES

Software Engineering & Architecture: Clean Architecture, Domain-Driven API design, fault-tolerant backends, circuit breakers, and automated testing.
Vector Search & Retrieval: ChromaDB, HNSW indexing, embedding similarity matrices, semantic search pipelines, and PDF document chunking.
AI Application Security: Intercepting LLM prompt injection attacks, out-of-distribution (OOD) clustering, and fail-safe gateway design.
Token Engineering & Performance: Byte-Pair Encoding (BPE) inspection, tiktoken compression, latency profiling, and LLM inference ROI analysis.
Standardized Protocols: Model Context Protocol (MCP) development for structured tool interoperability.

// 02 • TECHNICAL STACK & TOOLING

LANGUAGES & RUNTIMES

AI & APPLIED MACHINE LEARNING

SOFTWARE ARCHITECTURE & PROTOCOLS

// 03 • PRODUCTION SOFTWARE & REPOSITORIES

vector-guardrail

Problem Solved: Intercepting and neutralizing LLM prompt injection and jailbreak attacks at the API gateway layer before untrusted inputs reach downstream foundation models.

Core Stack: Python, Google Gemini, ChromaDB HNSW Indexing, SentenceTransformers, FastAPI
Software Architecture: High-speed vector embeddings similarity verification, Out-of-Distribution (OOD) clustering, and circuit-breaker resilience to guarantee zero system downtime during external API timeouts.

token-lab

Problem Solved: High inference costs and latency in generative AI applications caused by unoptimized prompt schemas and inefficient tokenization.

Core Stack: JavaScript, Node.js, Tiktoken, BPE Encoding Engines
Software Architecture: Interactive web laboratory providing real-time Byte-Pair Encoding (BPE) inspection across GPT-4o, Llama 3, and Qwen 2.5, schema compression benchmarking, and token ROI calculations.

vector-doc-engine

Problem Solved: Unreliable text extraction and slow semantic search across large PDF document corpora in enterprise RAG workflows.

Core Stack: TypeScript, ChromaDB, Google Gemini API, Document Processing
Software Architecture: Deterministic boundary-aware chunking pipeline, fault-tolerant vector generation, and sub-millisecond HNSW vector indexing.

zero-shot-vector-classification-eval

Problem Solved: High latency and model drift in text classification when using heavyweight fine-tuned models.

Core Stack: TypeScript, Google Gemini API, In-Memory Matrix Engine, L1 LRU Cache
Software Architecture: Zero-shot semantic clustering using pairwise cosine distance matrices with sub-millisecond L1 in-memory caching and out-of-distribution guardrails.

Official-Taiho-Toyota-Bearing-Catalogue-MCP-Server

Problem Solved: Bridging legacy engineering catalogue data with modern AI agents through standardized protocol interfaces.

Core Stack: Model Context Protocol (MCP), TypeScript, Structured JSON Schema
Software Architecture: Authoritative MCP tool provider delivering deterministic technical specifications and structured bearing queries directly to LLMs.

"Preoccupied with a single leaf... you won't see the tree.
Preoccupied with a single tree... you'll miss the entire forest."

— Miyamoto Musashi // Vagabond

Navigation Menu