AI / GenAI Engineer · Dubai, UAE — I build LLM agents and explainable decision systems for real businesses, and I implement the research papers behind them from scratch.
Open to AI / GenAI Engineer roles — UAE or remote.
| Project | Why it's worth a look |
|---|---|
| data-refinery | LangGraph pipeline that turns messy mixed-format files into governed, PII-safe, AI-ready datasets — DuckDB warehouse, lineage, quarantine. 34 offline tests + CI, MIT-licensed. |
| agentic-support-desk | Bilingual (Arabic/English) support-ticket triage split across four scoped agents — language, classification, response, escalation — each with a visible reasoning trace, so a support lead can see why a ticket was auto-answered or escalated. Zero-dependency: Python stdlib only, MIT-licensed. |
| aml-risk-agent-bedrock | Explainable transaction-risk and AML-flagging agent on AWS Bedrock: every flag names the rule that fired, the amounts and thresholds involved, and a SAR-ready narrative an analyst can file. Deterministic rules score; the LLM only writes the explanation. 37 tests + CI. |
| quant-agent | Adaptive, self-supervising paper-trading crypto system with a 9-layer risk pipeline. 81 tests + walk-forward backtest — and honest in the README about what isn't yet statistically significant. |
| langgraph-fraud-detection | Tiered fraud-detection agent: rule engine, anomaly scoring, and a Claude analyst for gray-zone transactions. |
| axxion-aios-orchestrator | Reference implementation of an agentic claims operating system — scoped agents, grounded verification, risk-tiered human gate, hash-chained audit trail. |
These are private while in active client conversations — happy to walk through the code or share access. Email malikk.monis@gmail.com for details.
- rasid — UAE Emiratisation quota tracker: reads an HR roster export, computes every expiring date/amount in plain Python, and holds any drafted message that states a figure it wasn't given.
- majlis — Bilingual (Arabic/English/Arabizi) support triage for UAE queues — redacts personal data before any model call and never auto-answers a regulator escalation.
- vouch — Answers security questionnaires from what you can actually prove — adversarial drafting with a deterministic guard against over-claiming.
I research a company's real, evidenced problem — then build a working prototype slice of the solution:
| Project | What it does |
|---|---|
| careem-support-resolver | AI agent that resolves Careem's #1 customer complaint (failed support) in seconds — grounded in policy + order facts, with safety guardrails. Backed by analysis of 400 real reviews |
| tabby-settlement-reconciler | Diffs a merchant ledger against Tabby's BNPL settlement report — catches missing payouts and fee discrepancies, ranked by money at risk |
| klaim-denial-shield | Explainable pre-submission claim denial-risk scorer + grounded appeal agent for UAE health insurance (DHA/DOH rejection codes) |
| vehicle-claims-demo | ClearPath Claims — full-stack demo of connected UAE motor insurance claims and underwriting pipelines sharing one knowledge base: agent clusters, red-team verification, risk-tiered human gate, hash-chained audit log |
| alaan-expense-agent | Explainable, FTA-defensible expense auto-decision engine |
| mamo-fairhold | Stress-tested rolling-reserve & hold engine for a payment facilitator |
| ziina-fairhold | Bounded, dated holds instead of full-account freezes for a licensed UAE payment company — holds the computed exposure, not 32x the dispute |
| keyper-tenant-risk | Explainable tenant default-risk underwriting for rent-now-pay-later |
| lune-transaction-enricher | Turns messy bank descriptors into structured merchant + category data |
| axxion-repair-audit | Explainable pre-payout repair-estimate leakage scorer for UAE motor claims |
| ammar | Tower maintenance dispatch: triages the ticket queue, matches technicians, and reopens any drafted closure that cites a record it wasn't given. 34 tests, hermetic CI. |
| axxion-aios-orchestrator | Runnable reference implementation of an agentic claims operating system — intake, routing, scoped agents, grounded verification, risk-tiered human gate, hash-chained audit trail |
| uae-startup-solutions | Ten researched problems across UAE startups (Stake, Huspy, Bayzat, Sarwa, Mal, and more) each solved with a compact explainable engine — CBUAE mortgage pre-underwriting, AAOIFI Shariah screening, grounded claims agents |
| motor-issuance-two-minute-bind | Agent-assisted motor-insurance application flow: a clickable eight-screen prototype and a timeline putting a bound policy at ~1:14, with an honest P50/P90 account of where UAE PASS and 3DS blow the budget |
| property-finder-trust-score | Explainable Listing Trust Score for Dubai real estate — perceptual-hash duplicate detection, bait-price outlier tests, CLIP photo-description mismatch; ships as CLI, API, and dashboard |
Note: these are independent, unaffiliated concept prototypes — not official products of the companies named, and built only on public information.
ai-employee-guard — a safety layer for AI agents given real authority: only asserts facts on record, only executes allowlisted actions, and hands off to a human when nothing on file resolves the request. Full audit trail. · shariah-guard — a compliance layer for AI-native financial products: grounds every Shariah-relevant AI output and escalates to a human board instead of improvising
langgraph-fraud-detection — tiered fraud-detection agent: rule engine, anomaly scoring, and a Claude analyst for gray-zone transactions · data-refinery — turns messy mixed-format files into governed, PII-safe, AI-ready datasets, with a DuckDB warehouse, local vector index, quarantine, and lineage · extended-stay-booking-agent — conversational booking for 30-365 night furnished stays, built two-brain so the LLM handles the conversation and never touches a price
RAG: rag (Lewis et al., 2020) · self_rag · corrective_rag · hyde — also rag-langchain-demo, the same pattern built on LangChain + Claude
Agents & reasoning: react_agent · reflexion · rewoo · pal · self_consistency · chain_of_verification · mixture_of_agents · dev_team (multi-agent SWE team)
Algorithms: hyperloglog (Flajolet et al., 2007) · minhash_lsh (MinHash + LSH banding)
prompt-guard — prompt-injection detection · agent-watchdog — loop detector + cost kill switch for agents · prompt-sentinel — git-native prompt regression testing · migration-guard · stripe-reconciler · api-breakcheck · env-sentinel · prompt-forge — expands a rough task description into a senior-engineer-quality prompt · taskforge — agent backend that executes a plain-English task with real tools (HTTP, file I/O, shell); FastAPI + Docker
Kashmiri-tts-model — text-to-speech for Kashmiri, a low-resource language
bioscan-ar — real-time anatomical AR mirror: webcam pose tracking drives a procedural render of a full skeleton, organs with a beating heart on a real cardiac cycle, and the arterial/venous trees down to peripheral branches
