ExperienceExperience

Part of AML frameworks team. Focused on ML infra which is the core part of recommendation, ads and search direction. Optimised Sparse MoE by +23.3%, increasing throughput of 8B fine ranking model by +15.36%. Optimised flash_attn_varlen_backward, increasing throughput by +4.84% for the same model.

Results

Sparse MoE optimisation

+23.3%

Throughput, 8B fine-ranking model

+15.36%

Throughput from flash_attn_varlen_backward

+4.84%

Part of Facebook Reels AI Foundation Team. Form the core AI logic of large-scale recommendation. Dabbled with ML training and experiments for different expert and foundational models. Improved Attention Kernels (fwd and bwd sdpa kernels) by > 33% resulting in training improvement > 5% qps. Implemented blockwise optimisation of SwigLU on MoE, resulting in 30% efficiency gain for SwigLU MoE, improving inference qps of 16B OneFlow model by 20% from 1490 QPS to 1790 QPS.

Results

Faster SDPA kernels, fwd + bwd

>33%

Training QPS

>5%

SwiGLU MoE efficiency gain

30%

Inference QPS, 16B model (+20%)

1490 → 1790

Built an Agentic RAG tool that automated 30% of SQL query construction for post-trade regulatory reporting. Remodelled the LLM Suite API with FastAPI, driving adoption across 5 global teams, and engineered metadata-driven vector DB switching that cut p95 retrieval latency from 6s to 2s.

Results

Of SQL query construction automated

30%

p95 retrieval latency

6s → 2s

Global teams adopting the API

5

Designed a PEFT fine-tuning pipeline on AWS EC2 for TinyLlama, StabilityAI, and Qwen models. Achieved a perplexity drop from 16 to 4 for Java Card code completion, with the fine-tuned model passing 90% of regex test cases versus 0% at baseline.

Results

Perplexity, Java Card completion

16 → 4

Regex tests passed vs baseline

90% vs 0%

Re-architected a HoloLens VR app, migrating from Google Cloud to OpenAI and Anthropic and adding RAG-powered memory using gaze and pointer data for richer query relevance. Built Milvus vector endpoints that improved device energy efficiency by 20% over Pinecone, and containerised the Python server with Docker.

Results

Device energy efficiency vs Pinecone

20%

Led end-to-end development of an internal Customer Management System with Angular and Java Spring Boot, cutting decision times for 50+ sales users. Optimised contract operations by 150% via Java concurrency and delivered a data migration workflow within 2 days.

Results

Contract operations optimised

150%

Sales users with faster decisions

50+

Data migration workflow delivered

2 days