AI/LLM Engineer · Forward Deployed Engineer

Almaty, Kazakhstan · sabit-shaikholla.github.io · LinkedIn · saba.shaiholla@gmail.com

On the work side I've built enterprise RAG over internal knowledge: 100+ code repositories, 1000+ Confluence pages and 5000+ internal documents, most of it scanned PDFs and badly formatted DOCX where top-k retrieval breaks down. The pipeline runs an automated RAGAS evaluation loop (faithfulness 0.89, answer relevance 0.92 on a 15K-query benchmark), and for the messy-document part I use agents that decide what to read themselves instead of trusting nearest-neighbour search.

On the serving side: deployed Kimi-K2.6 on-premise with SGLang (MoE, 100+ tokens/sec throughput, 52% lower TTFT than baseline) and GPT-OSS 120B with vLLM. Also built a marketplace for AI agent skills on APM (Agent Package Manager) — skills as installable packages with versions and dependency resolution, instead of being copied between projects by hand.

Background: Spring Boot microservices, event-driven backends, edge/IoT systems, plus a stretch leading a team of 10 on platforms with 3M+ MAU. A lot of my work is customer-facing — working out what they actually need before anything gets built.

Code here, reasoning on my site.

• Agentic document search on Google ADK + Gemini 3 Flash: the agent navigates files with tools instead of an embedding index.
• Evaluation in the loop: OpenEvals graders as agent reflection steps, groundedness scoring, self-correcting retrieval. Posts on both below.

An agent that explores a document folder itself rather than trusting nearest-neighbour search on pre-embedded chunks. It runs a Scan → Deep Dive → Backtrack loop over nine filesystem and vector tools (scan_folder, grep_search, semantic_search, parse_file, …), so it can follow references and retry when retrieval misses something.

Google ADK · Gemini 3 Flash · Docling · pgvector / DuckDB · FastAPI

RAG pipeline that grades every retrieved document for relevance before using it. The score routes the agent into Correct / Incorrect / Ambiguous branches, falling back to web search if nothing passes the threshold. OpenEvals graders run inside the agent loop as reflection steps, not just offline tests.

LangGraph · Gemini 2.5 Flash · Ollama (Qwen 2.5 7B) · OpenEvals · Tavily

Managed RAG end to end: upload a PDF, get answers with grounding metadata pointing at source chunks. Runs on an Oracle Cloud always-free instance behind Nginx + Let's Encrypt, with CI pushing releases. Setup documented step by step.

Gemini File Search API · Streamlit · Nginx · GitHub Actions

Marketplace for AI agent skills built on APM (Agent Package Manager): skills as packages you can install, with versions and dependency resolution. Not public, no repo to link.

Agent Package Manager · Python

Selenium scraper that gets past FBREF's 403s, 584 engineered features across 17 statistical categories, six seasons of matches. Reaches 61.2% test accuracy on three-way Win/Draw/Loss against a 33.3% baseline. The scraper broke after OPTA removed the advanced tables in Jan 2026, but the datasets are in the repo, so the modelling still runs.

Python · scikit-learn · Selenium · pandas

• Beyond RAG: Building an AI Agent That Explores Documents Like a Human Researcher _{Feb 2026}
• How I Setup Oracle Cloud Always Free Instance for Gemini API File Search Tool _{Nov 2025}
• Revolutionizing RAG: Why Gemini File Search Tool is the great RAG-as-a-Service _{Nov 2025}
• Enhancing LLM Agent Reliability with Corrective RAG and OpenEvals _{May 2025}

→ All posts · TIL notes · RSS

_{Open to AI/LLM and forward-deployed engineering roles. Email is fastest: saba.shaiholla@gmail.com.}