SC\OSPby sathiska priyad

AI research & engineering for enterprise systems.

Lead AI Engineer specializing in Retrieval-Augmented Generation, distributed backend systems, and strict schema validation for high-stakes enterprise environments.

Technical Arsenal

AI, Systems & Observability

Core AI & Orchestration

PythonLangChainLangGraphLlamaIndexOpenAIAnthropic

Observability & Evals

LangSmithRagasLLM-as-a-JudgePromptLayerDatadog

Fine-Tuning & Inference

PyTorchHugging FaceLoRA / QLoRAvLLMOllamaTRL

Vector Search & Data

PineconeFAISSQdrantpgvectorPydantic

Distributed Systems

AWSDockerKubernetesFastAPIREST / GraphQL

Frontend Engineering

TypeScriptReact.jsNext.jsTailwind CSSKnockout.js

Professional Experience

BloomX Analytica

Lead AI Engineer
05/2025 - PresentLondon Area, UK
  • Defined the end-to-end AI technical roadmap and architected production-grade LLM/RAG pipelines using Agentic AI patterns.
  • Led foundation model evaluation and authored company-wide responsible AI standards covering data privacy and fairness.
PythonLangGraphLLMsRAGAgentic AI

B&S

Associate Technical Lead (Frontend)
08/2022 - PresentThe Randstad, Netherlands
  • Owned the frontend architecture for a major European e-commerce platform, driving full-stack performance overhauls and API optimizations.
  • Pioneered the company's first LLM-powered semantic search and shipped complex production-grade React.js UIs.
React.jsGraphQLMySQLAdobe CommerceMagento 2Knockout.js

Wiley Sri Lanka

Senior Software Engineer
10/2021 - 10/2022Colombo, Sri Lanka
  • Core engineer on a global e-learning LMS platform, building scalable full-stack features using PHP and Django.
  • Acted as a key technical resource for distributed cross-functional squads, ensuring high-quality, well-tested code delivery.
PHPDjangoFull-StackLMS

M.SaaS

Senior Software Engineer
05/2021 - 10/2021Colombo, Sri Lanka
  • Led frontend development for the company's core B2B SaaS product, building performant UI modules using React.js.
  • Collaborated with product teams to translate design specs into accessible interfaces and introduced component-level unit testing.
React.jsHTML5CSSSaaS

Open Source

Profile: @spriyads-vaultView GitHub

adaptive-rag-engine

Python

Cloud-Ready AI Architecture: Engineered a FastAPI-based retrieval pipeline integrating FAISS and BM25, designed for seamless deployment onto AWS EC2/ECS instances to deliver scalable LLM grounding.

View Repository

langchain-multi-agent-workflow

Python

Agentic Orchestration Framework: Architected a multi-agent AI system capable of decomposing complex user intents into specialized sub-tasks, mimicking a specialized cross-functional human team.

View Repository

customer-support-agent

Python

Autonomous Support Agent: Developed an intelligent customer support agent integrated with LangSmith for enterprise-grade observability, trace debugging, and continuous performance monitoring.

View Repository

@spriyad

Hugging Face

Exploring and deploying open-source LLMs, embedding models, and custom fine-tunes for enterprise RAG pipelines and agentic workflows.

Model Fine-tuningEmbeddingsOpen Source AI
View Profile

Pragmatic RAG: My Architectural Blueprint

Engineering for context and truth. This is my proven methodology for building enterprise-grade Retrieval-Augmented Generation systems that eliminate hallucinations and scale securely, following strict industry best practices.

System Architecture Flow
1. Raw Data Sources
PDFs, SQL, Confluence, APIs
2. Semantic Processing
Structure-Aware Parsing & Chunking
3. Hybrid Retrieval Engine
Dense Vectors + BM25 Sparse
4. Query Orchestration
Intent Routing & Expansion
5. Grounded Generation
Strict Schema Injection
01

Data Ingestion & Parsing

Most RAG fails at the PDF parser, not the LLM. I build robust extraction pipelines that preserve document structure—tables, headers, and metadata—before chunking even begins. If context is lost here, the vector database cannot save it.

02

Semantic Chunking & Embedding

Static token limits destroy meaning. I implement semantic chunking strategies that respect natural breakpoints in the data. Combined with high-dimensional embeddings, this ensures retrieval grabs complete thoughts, not broken fragments.

03

Vector Database Architecture

I design isolated namespaces and hybrid search indices (dense vector plus sparse BM25) to guarantee low-latency, high-precision retrieval across massive, unstructured enterprise datasets.

04

Query Rewriting & Routing

Users rarely ask perfectly formulated queries. I place a lightweight routing and query-expansion layer before the database to translate messy human intent into optimized semantic searches, drastically improving recall.

05

Context Injection & Generation

The final prompt is a strict schema. I enforce hard boundaries between the retrieved facts and the system instructions to prevent prompt injection and force the target LLM to synthesize answers strictly grounded in the retrieved truth.

Surgical Fine-Tuning: Maximum Impact, Minimum Compute

My methodology for adapting open-weights models. By leveraging Parameter-Efficient Fine-Tuning (PEFT) and rigorous data curation, I deliver specialized AI capabilities that outperform massive generic models at a fraction of the inference cost.

Performance Benchmark: Base vs Fine-Tuned

* Illustrative metrics demonstrating the typical ROI of targeted LoRA adapters on domain-specific tasks compared to zero-shot prompting on base 8B-70B parameter models.

01 //

Dataset Curation & Formatting

Data quality beats volume every time. I spend the majority of fine-tuning cycles deduplicating, filtering noise, and formatting golden conversational pairs. A thousand perfect examples will always out-train a million mediocre ones.

02 //

Base Model Selection

I benchmark open-weights models against the specific domain task, balancing the required reasoning capabilities against the actual hardware budget and latency requirements of the target deployment environment.

03 //

Parameter-Efficient Fine-Tuning

I use LoRA and QLoRA to train targeted adapters on specific behaviors without catastrophically forgetting the base model's general knowledge. This allows for rapid experimentation and swappable domain expertise at runtime.

04 //

Evaluation & Benchmarking

Standard benchmarks rarely reflect production reality. I build automated LLM-as-a-judge pipelines and domain-specific regression tests to ensure the model solves the business problem without degrading baseline safety guardrails.

05 //

Deployment & Inference Optimization

An accurate model is useless if it is too slow. I deploy using continuous batching and KV-cache optimization engines to maximize GPU utilization and slash Time-To-First-Token.

System Architecture & MLOps Ledger

I don't just build models on my laptop; I deploy them securely to millions of users. This is the production-grade stack I use to manage the entire machine learning lifecycle at scale.

Data Ingestion & Processing

Apache AirflowSnowflakeApache Kafkadbt

High-throughput, fault-tolerant pipelines for structured and unstructured enterprise data.

Training Infrastructure

RayAWS SageMakerPyTorch DistributedNVIDIA DGX

Distributed multi-node cluster orchestration for large-scale model pre-training and fine-tuning.

Model Registry & Tracking

MLflowWeights & BiasesHugging Face HubDVC

Strict version control, experiment tracking, and artifact management for reproducible AI.

Serving & Inference

vLLMTensorRT-LLMKubernetesTriton Inference Server

Ultra-low latency, high-throughput model serving with continuous batching and KV-cache optimization.

Business Impact & Engineering Leadership

I build AI that makes money, saves time, and elevates the developers around me. A Lead Engineer bridges the gap between the engineering team and the executives, focusing on measurable ROI and strategic execution.

$1.2M
ARR Saved

Cloud Compute Optimization

Reduced cloud compute costs by 45% ($1.2M ARR) by migrating from proprietary LLM APIs to self-hosted, quantized Llama-3 models via vLLM on spot instances, without degrading response quality.

32%
Precision Lift

Search & Retrieval Accuracy

Increased retrieval precision by 32% and downstream user conversion by 18% through the implementation of a hybrid dense-sparse RAG architecture with cross-encoder reranking.

6+
Engineers Mentored

Team Leadership & CI/CD

Mentored a team of 6 junior-to-mid level engineers, establishing strict CI/CD pipelines for ML models and reducing deployment failure rates from 15% to near-zero.

0 to 1
Product Delivery

Strategic Execution

Bridged the gap between executive product vision and engineering execution, translating ambiguous business requirements into deterministic, scalable AI system architectures delivered ahead of schedule.

Professional Endorsements

Verified LinkedIn Recommendations

View on LinkedIn

"One of the most reliable and skilled front-end developers I've worked with."

I've had the pleasure of working closely with Sathiska on several projects. Sathiska played a key role in building a strong technical foundation that allowed us to onboard many new businesses smoothly and efficiently. Beyond their technical strengths, what really stood out was Sathiska's dedication, excellent communication, and team-first attitude.

Daniël Laan

Development Manager IT - E-commerce
@ B&S InternationalNetherlands

"An intelligent and experienced software engineer."

Sathishka is a dedicated teammate and excellent frontend and UI engineer. He takes up challenges with much positive attitude and always help his other teammates. It was a pleasure to work with him.

Buddhini Perera

Program Manager
@ Digital TransformationSingapore

"I was really impressed with his quality of work and attitude."

Sathishka Is always keen to learn and take challenges. His eagerness to learn new things has helped him master quite a few technologies quickly. He has been a great team player and has been trustworthy. I would highly recommend him for any organization.

Waruna Viraj Perera

Senior Front-End Developer
@ Wunderman ThompsonUnited Kingdom
Live Production Showcase

Data Dictionary & Compliance Mapper

Launch Application

Project Overview

A live, deployed instance of a schema intelligence engine built to solve a critical enterprise bottleneck: identifying PII and mapping compliance rules across massive, undocumented SQL databases. Upload a raw DDL file or use the sample schema to watch the model process, structure, and tag the data in real-time.

Architecture Highlights

PII Detection Engine

Real-time schema parsing to automatically detect and classify Personally Identifiable Information.

Regulatory Mapping

Aligns raw database structures with GDPR and CCPA regulatory requirements dynamically.

Zero-Trust Processing

Secure, client-side processing pipeline ensures sensitive DDL schemas never leave the browser.

Tech Stack

Next.jsTailwindLangChainEnterprise RAG
compliance-mapper.run.app