Software Engineer · AI Systems
Murad Al-Balushi
I build the control, evaluation, and safety layers that make LLM systems reliable enough for production.
- ✓Replaced subjective model evaluation with execution-based validation
- ✓Enforced LLM spend limits in agent workflows
- ✓Constrained model outputs to eliminate hallucination
- ✓Built compliance-ready fintech infrastructure — zero security incidents
Professional Experience
360Remit
Software Developer
Muscat, Oman
Jan 2025 – Mar 2026
- Owned end-to-end VAPT for a regulated fintech platform as the risk authority between vendors and engineering; validated findings, cut false positives 40%+, closed 100% of critical issues pre-launch, zero security incidents at go-live.
- Engineered vendor synchronization pipeline for 500k+ records (delta detection, conflict resolution, bidirectional sync), cutting manual processing from 3–5 days to under 5 minutes with 100% DB integrity.
- Delivered MTO, eKYC, and AML integrations and designed phased infra (DR, capacity, data residency), enabling platform launch within regulatory deadlines while unblocking user-facing onboarding flows.
- Built SQL/Python humanization pipeline converting 500k+ vendor records to presentation-ready data, eliminating manual cleaning and cutting prep time 95%+.
Highlight Projects
Evaluation, cost control, safety, and research reproduction — the layers that make AI systems production-ready
AI Code Generation Evaluation Engine (Code Arbiter)
Execution-based benchmarking — run the code, classify the failure
Replaced subjective LLM code review with deterministic execution-based validation. Runs generated code in isolated Docker sandboxes, classifies failures across syntax, runtime, logic, and temporal reasoning, and benchmarks multiple models under identical conditions.
CostPlan – LLM Cost Enforcement Proxy
Open-source circuit breaker for autonomous agent API spend
Built an open-source transparent proxy that enforces per-call and per-session budget limits on LLM API calls, with cache-aware pricing and zero-latency SSE streaming — preventing unbounded spend in autonomous agent workflows.
Production AI Support Agent (Guardrail-First)
Risk-aware LLM-powered support agent reducing customer support load
Deployed a guardrail-first AI support agent handling live customer tickets with Stripe-backed context and deterministic escalation logic, designed to fail safely under uncertainty in a production SaaS environment.
Brain2Qwerty v1 — MEG Brain-to-Text Reproduction
Reproduced Meta FAIR's non-invasive brain-to-text decoder on public MEG data
Reproduced Meta FAIR's Brain2Qwerty v1 — a 623M-param Conv+Transformer that decodes typed sentences from MEG brain recordings — on the public SpanishBCBL dataset, matching the documented no-LM target of ~0.38 character error rate across 19 subjects. Full pipeline on a single GCP L4 GPU: gated 250GB dataset handling, a ~1.5-day training run, and a CUDA OOM diagnosis fixed with zero fidelity cost.
Technical Skills
Technologies I use to build, ship, and evaluate production systems
Languages
AI & LLM
Backend
Infrastructure
Databases
Tools & Integrations
Let's Connect
Open to opportunities, collaborations, and interesting problems.