About
I'm Arnol P S, Senior Engineer, Data Science at Reflections Info Systems, based in Kerala, India. In practice the work is forward deployed: I work on site with clients and take generative AI from the first discovery session to deployment on their own on-premise servers, building on 5+ years across data science, NLP and computer vision.
My work has included multi-agent systems, hybrid RAG pipelines that combine vector and graph databases, and real-time conversational AI. It runs from chunking and embedding pipelines to LLM guardrails and observability.
Most recently I spent September and October 2026 on site with a government client where, working in a team of three, I took on most of the technical delivery of an air-gapped, bilingual English/Arabic assistant. It was deployed on the client's on-premise servers and handed over for user acceptance testing on 28 October 2026. Read the case study.
What forward deployed means to me
How I work
-
Start with discovery and write the requirements down
Before anything gets designed, I run discovery sessions with the people who will use the system. Then I turn what they told me into written requirements the client can review and correct. When a scope question comes up later, we settle it from the document.
-
See it through on the client's servers
I run the releases on the client's servers and stay to fix what only shows up there. On my last engagement that meant fixing Arabic PDF text that came out with its letters reversed, and resyncing a server clock before it affected sign-in.
-
Measure under load before claiming anything
Before I say a system can handle something, I load-test it on the same class of hardware first, then again on the client's own servers once it is deployed. A number goes into a report only with the conditions it was measured under, so someone else can repeat the test. On my last engagement I ran the GPU benchmarks myself, on the client's hardware.
-
Document for the people who run it after handover
After handover, the client's own team deploys, checks and rolls back the system on its own. I write for them: the steps in the order they happen, and the check that shows each step worked.
Toolbox
Client delivery
- Discovery
- Requirements (BRD)
- Solution design (SDD, LLD)
- On-premise and air-gapped deployment
- UAT support
- Handover documentation
LLM serving and infrastructure
- vLLM
- Tensor parallelism
- FP8 KV-cache quantization
- Triton Inference Server
- Text Embeddings Inference
- Speech AI (ASR, TTS)
- NVIDIA H200 GPUs
- CUDA
- GPU capacity sizing
- Load testing (Locust)
- Docker
- Kubernetes
- AWS (Bedrock, Lambda, API Gateway)
RAG and agents
- Retrieval-augmented generation (RAG)
- GraphRAG
- Hybrid search (BM25, SPLADE)
- Cross-encoder reranking
- Reciprocal Rank Fusion
- Qdrant
- Neo4j
- Multi-agent systems
- LangGraph
- LangChain
- Model Context Protocol (MCP)
- Document AI (Docling, OCR)
- Multilingual and English/Arabic search
- Claude, Gemini, Qwen
Safety, governance and observability
- LLM guardrails
- Prompt-injection defense
- PII detection
- Qwen3Guard
- Llama Guard
- Presidio
- Keycloak
- ISO 42001
- Langfuse
- Prometheus and Grafana
- RAGAS
Machine learning and computer vision
- PyTorch
- Transformers
- Fine-tuning (LoRA, QLoRA)
- Scikit-learn
- ONNX Runtime
- OpenCV
- YOLO
- DINOv2
- Vision-language models
Engineering
- Python
- TypeScript
- JavaScript
- React
- SQL
- FastAPI
- Pydantic
- WebSockets
- PostgreSQL
- Redis
- Elasticsearch
- Claude Code
Education
-
Master of Science in Computer Science (Data Analytics)
Indian Institute of Information Technology and Management-Kerala (IIITM-K)
Cochin University of Science and Technology
First Class, CGPA 7.60
-
Bachelor of Science (Computer Science, Mathematics, Statistics)
Kristu Jayanti College (Autonomous), Bengaluru
Bangalore University
First Class
Certifications
Claude Certified Architect - Professional
Issued by Anthropic in August 2026.
Verify on Credly (opens in a new tab)-
Google Data Analytics Professional Certificate
Google
-
Building Real-Time Video AI Applications
NVIDIA Deep Learning Institute
-
Getting Started with Deep Learning
NVIDIA Deep Learning Institute
Recognition
Shout-Out award
Received for my contributions to the AI Management System documentation and to preparedness for the ISO 42001:2023 surveillance audit.
Languages
-
English
Professional
-
Malayalam
Native
-
Swedish
Beginner