Dubai, UAEOpen to new roles — Dubai or remote

Ahmad Rishi

Senior AI / LLM Engineer · AI Team Lead

GenAI · Agentic Systems (LangGraph, MCP) · RAG · LLMOps · Python · AWS

I build production LLM systems — agentic workflows, RAG, and the evaluation harnesses that prove they work — and I own them from the first commit to production.

View workRésuméGitHub Email
Blind eval: my RAG vs the incumbent tool
240–30–13Blind eval: my RAG vs the incumbent tool283 real questions, LLM judge
Sole-authored agent definitions
38K linesSole-authored agent definitions16 of them, each blame-verified at 100%
Engineering, ~2 in production GenAI
4 yearsEngineering, ~2 in production GenAISoftware Engineer → AI Team Lead in ~16 months

Receipts

What the numbers actually count

Every figure below was measured in the source repository it came from — no estimates, no projections, no round numbers.

Questions in the blind eval corpus

283

Questions in the blind eval corpus

Pairwise, my RAG against the incumbent. An LLM judge and a human domain expert agreed 91.5% of the time.

Agentic platform I initiated and lead

~116K LOC

Agentic platform I initiated and lead

I wrote the root commit and merged 94 feature branches. It runs 16 workflow types over a 46-table MySQL schema; the LOC figure is the platform's size, across ~8 contributors.

Lines of agent definitions, sole author

38K

Lines of agent definitions, sole author

16 model-specialised definitions — discovery, review, test-writing — each verified at 100% by git blame.

Pass-rate swing from changing only the judge

37.4 pts

Pass-rate swing from changing only the judge

82.7% → 45.3% on byte-identical outputs, against ±9–11 sampling error. 37.6% of verdicts failed to parse.

Cheaper per token after a model migration

~12×

Cheaper per token after a model migration

Gated on an LLM-as-judge replay harness of real production conversations, not on the price tag.

Sustained ceiling, found by load-testing to failure

~65 req/s

Sustained ceiling, found by load-testing to failure

30,300 requests ramped 26→940/s. Flat throughput past the ceiling proved a saturated worker pool, not the network.

Selected work

The systems I built and lead

These run in production or ran long enough to be measured. Each card carries the problem, the numbers it moved, and what I owned on it — with the public repository one click away.

All projects 
Lead case study2025 – 2026

Agentic Development Board

Turning a business request into shipped code meant a chain of manual handoffs — speccing, decomposition, implementation, tests, documentation — with no single place to answer where a task actually is.

Feature branches merged
94Feature branches merged
Workflow types
16Workflow types
MySQL schema
46 tablesMySQL schema
Project size
~116K LOCProject size

What I owned — Initiated the platform and led it as maintainer; personally built the LLM provider abstraction, the subprocess layer, the Atlassian auth layer and the self-healing API-generation workflow.

  • Python
  • FastAPI
  • LangGraph
  • Kafka (KRaft)
  • MySQL
  • Next.js 15
  • React 19
  • Terraform
Repo on GitHub 
  • 2025 – 2026

    AI Agent Kits & Spec Skills

    Coding agents produce plausible code that ignores a codebase's real conventions, and every team member was re-teaching the same context to the same models by hand.

    Agent definitions
    16Agent definitions
    Lines, sole author
    38KLines, sole author
    Blame-verified
    100%Blame-verified

    What I owned — Sole author of all 16 agent definitions and the conventions behind them.

    • Claude Code SDK
    • Cursor CLI
    • MCP
    • Python
    • Markdown
    Repo 
  • 2024 – 2025

    Enterprise Knowledge RAG

    Business-flow knowledge lived in Jira and Confluence, so answering a routine process question meant interrupting an engineer who happened to remember the ticket.

    Blind pairwise, LLM judge
    240–30–13Blind pairwise, LLM judge
    Human expert judge
    246–34–3Human expert judge
    Judge agreement
    91.5%Judge agreement
    Jira issues indexed
    30–60KJira issues indexed

    What I owned — Built the retrieval pipeline, the citation construction, the async serving layer and the blind evaluation that cleared it for production.

    • Python
    • LangChain
    • LangGraph
    • Pinecone
    • Flask (async)
    Repo 
  • 2025 – 2026

    CodeHarbor

    Reviewing every merge diff across an ERP estate for security defects is work no human team sustains, and a plain LLM reviewer produces confident findings it cannot substantiate.

    Commits
    207Commits
    Insertions
    ~25KInsertions
    Security checks
    9Security checks

    What I owned — Initiated the repo and led the scan engine — sandboxed execution, the 9-check checklist, evidence gating and de-duplication.

    • Python
    • FastAPI
    • LangGraph
    • Claude Code SDK
    • Pinecone
    Repo 
  • 2025

    LLM Evaluation Framework

    Deciding whether a cheaper model can replace an incumbent in production is usually settled by a demo, a benchmark table, or the price sheet — none of which replay what users actually send.

    Judge-choice swing
    37.4 pointsJudge-choice swing
    Verdicts failing to parse
    37.6%Verdicts failing to parse
    Cost of the candidate
    ~12× cheaper / tokenCost of the candidate

    What I owned — Designed and built the harness, the seven evaluation arms, the reciprocal cross-judging design and the self-preference probe.

    • Python
    • OpenAI
    • Anthropic
    • Gemini
    • DeepSeek
    Repo 

How I work

Four habits that survive an interview

Each of these came from a system that either shipped or broke in a way worth reporting. None of them is a preference I picked up from a blog post.

  • Evaluation

    I decide with evals, not vibes.

    Before a model reaches production it replays real conversations through a multi-provider harness — seven arms, deliberate reciprocal cross-judging, a self-preference probe. That is how a candidate ~12× cheaper per token cleared its migration, and how I found that swapping only the judge moved the pass rate 37.4 points on byte-identical outputs.

  • Agentic delivery

    Agents ship behind human approval gates.

    The internal platform I initiated drives a business request through spec → plan → code → tests → docs as LangGraph workflows, and a person signs off at every gate. Sixteen workflow types over a 46-table schema; one of them — the natural-language → secure-API workflow I wrote — carries a self-healing validate→repair loop that retries the model rather than the reviewer.

  • Retrieval

    Citations are structural, not requested.

    In the production Jira/Confluence RAG, citations are generated by construction — a citation cannot reference a document that was never retrieved, because there is no code path that would let it. Blind pairwise against the incumbent tool over 283 real questions: 240–30–13.

  • Proof

    I report the real ceiling, not the peak.

    I ramped the WhatsApp agent I built from 26 to 940 req/s across 30,300 requests until it fell over, then reported the ~65 req/s it actually sustains rather than the best number the run produced. Successful throughput stayed flat regardless of arrival rate — that flatness is what proved a saturated worker pool rather than a network limit, and it pointed remediation at concurrency. Every figure on this site was arrived at the same way.

Experience

Software Engineer to AI Team Lead in about sixteen months

Open any role for what it actually involved. The team I lead is four to six engineers, and I mentor the new-grad hires on it.

  1. May 2025 – Present

    Lead a 4–6 engineer AI team, own the agentic-platform roadmap and org-wide AI tooling, and mentor engineers and new-grad hires. Initiated and lead the Agentic Development Board — wrote its root commit, merged 94 feature branches to master, and drove it to 16 workflow types over a 46-table MySQL schema across a ~116K LOC platform. Authored 16 model-specialized agent definitions as drop-in agent kits — 38K lines, sole author, each blame-verified at 100%. Initiated CodeHarbor, an agentic LLM security scanner, and led its scan engine (207 commits, ~25K insertions). Stood up the LLMOps backbone (self-hosted LiteLLM gateway, Langfuse tracing) and migrated production traffic to a model ~12× cheaper per token, gated on an LLM-as-judge replay harness. Built a WhatsApp agent pilot on n8n — all 22 workflows and 357 nodes — and load-tested it to destruction: ~65 sustained req/s ceiling across 30,300 requests ramped 26→940/s.

  2. Sep 2024 – May 2025

  3. Sep 2024 – Apr 2025

  4. Jan 2024 – Sep 2024

  5. Jul 2022 – Mar 2024

About

Who is behind the commits

I am an AI/LLM engineer in Dubai and currently AI Team Lead at Maids.cc, where I lead a 4–6 engineer AI team and own the agentic-platform roadmap. Four years of engineering, roughly two of them building production generative AI; Software Engineer to AI Team Lead in about 16 months.

I initiated the Agentic Development Board — an internal platform that drives a business request through spec → plan → code → tests → docs as LangGraph multi-agent workflows behind human approval gates. I wrote its root commit and have led it as maintainer across ~8 contributors, and personally built the dual-engine LLM provider abstraction, the subprocess layer that makes forking a 600-second child safe, the Atlassian OAuth 2.0 + JWT session layer end to end, and the natural-language → security-validated API workflow with its self-healing validate→repair loop. It now runs 16 workflow types over a 46-table MySQL schema across a ~116K LOC codebase.

Before that I built the production Jira/Confluence RAG that was preferred over the incumbent tool 240–30–13 across 283 blind pairwise questions (246–34–3 judged by a human domain expert, the two agreeing 91.5% of the time), and the multi-provider evaluation harness that cleared a production migration to a model ~12× cheaper per token. Every number on this site was measured in the source repository it came from, and every system behind them is one I initiated, led or built.

Education
Bachelor of Science in Information TechnologySyrian Private University · 2019 – Jan 2024
Certifications
Associate AI Engineer for DevelopersDataCamp · Jan 2025
Developing AI ApplicationsDataCamp · Nov 2024
Languages
Arabic (native), English (fluent)

Stack

Generative AI & LLMs
RAGAgentic & multi-agent systemsPrompt engineeringStructured outputsTool / function callingMCPLLM-as-judgeNLP
Frameworks & Orchestration
LangChainLangGraphLangSmithn8nClaude Code SDKCursor CLIHugging Face Transformers
Models & Providers
Claude (Anthropic API)GPT / OpenAIGeminiDeepSeekLlamaEmbeddings
RAG & Vector DBs
PineconeMilvusChromaDBSemantic chunkingHybrid searchRerankingCitations
LLMOps & Evaluation
LiteLLM gatewayLangfuseMLOpsEvaluation harnessesCost optimizationTracing
Backend & Full-Stack
FastAPIFlaskRESTasyncioMicroservicesKafkaCeleryRedisSSE / WebSocketsPostgreSQLMySQLMongoDBNext.jsReactAngularTypeScriptNgRxTailwind
Cloud & DevOps
AWS (EC2, ASG, RDS, ALB, ECR, ACM, IAM, SSM, EKS, Bedrock)DockerKubernetesTerraformGitLab CI/CDGitHub Actions
ML / CV & Security
PyTorchTensorFlowOpenCVEasyOCRDE-GANLLM security (prompt-injection defense, OWASP LLM Top 10)Secrets scanningPlaywright E2E
Languages
Python (expert)SQLTypeScript / JavaScriptPHP

Contact

Tell me what you're building

I read everything that comes in. If you're hiring, the résumé and the repositories will answer most of it first.

Based in Dubai, UAE and open to remote. Currently leading a four-to-six engineer AI team at Maids.cc.