Muhammad Ibrahim

San Sebastián, Spain

• Owned the GLM-5.3 serving stack end to end within 24 hours of joining: from crashing and failing benchmarks to passing a 200k-token SLA (64 concurrent users, 100% success, TTFT P90 34 s vs a 70 s limit).
• Re-architected it as prefill/decode disaggregation with context-parallel prefill: same SLA on 5 nodes instead of 11, 65% more throughput, 3.6× per node.
• Found and fixed silent correctness and stall bugs, each validated on GPUs before rollout: LMCache KV corruption (gibberish outputs from up to 16% to 0 of 2,000), a lost-completion race, a missing-key read hang, a load-failure stall (max ITL 575 s → 14.7 s).
• Zero-downtime production rollouts; root-caused truncated streams and 503s to autoscaler terminations and Istio/Envoy keep-alive mismatches.
• Owned the company-wide GPU job platform (SkyPilot): 0 failed jobs, EFA networking for ~2× faster training, NVIDIA MIG partitioning as code.
• Secured $1.2M in contracts through live demos for AWS Accelerate; mentored 4 junior engineers (2 promoted within 6 months).

San Sebastián, Spain

• CompactifAI: led a team of 6 building a 70%-compressed LLM, adopted by 3 enterprise clients for edge deployment.
• Built an AWS Batch Monte Carlo simulation platform, cutting GPU costs by 25% ($50K/year).

Karachi, Pakistan

• Automated 80% of loan-processing workflows (1,200+ loans/month), cutting processing time by 40%.
• Built Java/Quarkus microservices handling 50K+ daily transactions with zero security incidents.

Karachi, Pakistan

• Built an Android RFID warehouse app and automated SAP HANA workflows (40% fewer data-entry errors).