I'm

Avrodeep Pal

★

Technologies

• LLM Serving & vLLM Migration: Delivered ~17x LLM throughput on staging by migrating serving to vLLM, after a controlled benchmark of 7 serving setups showed the serving software, not the GPU, was the bottleneck: 100 mixed requests dropped from 213s to 12.6s. 64 concurrent streams now finish in 2.2s where the old server froze past 40, with GPU memory budgeting enforced before any deploy and 59 of 59 live checks passing. The benchmark also surfaced 8 platform bugs, all fixed, and produced 13 ranked recommendations for leadership.

• DeepSeek V4 Flash Bring-Up: Scored 92.5/93 on a 10-phase acceptance suite across ~20,500 requests while bringing up a large mixture-of-experts model on staging with tensor and expert parallelism, extending the platform to manage multi-GPU replicas, backed by 463 unit tests. Recalled 48 of 48 planted facts in contexts up to ~260K tokens, handled 100 simultaneous calls in 18.25s, ran a 30-minute soak with zero restarts, and recovered from a killed GPU process on its own in 88s. Deploys dropped from an 8-minute reload to 3.6 minutes with the model left running.

• Security, Reliability & Deployment: Audited a ~114,000-line Python platform by tracing request paths end to end, then mapped failure modes across core subsystems, delivering ranked findings and phased fix plans for leadership. Hardened input and file handling with 80+ regression assertions and restored correct HTTP status codes so monitoring and retries work as intended. Reworked deployment tooling across 100+ files, removing ~573 duplicate lines, verified by 32 offline tests and parity across 123 commands; now in validation ahead of rollout.

• Image-to-Video Optimisation & GPU Debugging: Sped up image-to-video generation ~3.5x (398.7s to ~115s) by stacking and measuring each optimisation, including FP8 and torch.compile. Traced broken multi-GPU output to silent data corruption in direct GPU-to-GPU copies, reproduced with no model loaded (40 of 40 transfers corrupt, 0 of 40 through host memory), then fixed it by routing copies through the CPU. Also diagnosed thermal throttling and handed it to the hardware team with evidence.