Life sciences
Enterprise architecture and technology leadership in global pharma, where validation and change traceability are central.
Architect · AI on GPUs · Digital-asset systems
AI for regulated industries: self-hosted LLMs on H100, L40 and L40S, and document AI for financial and regulatory filings. Trading infrastructure: market-making bots in Go, a trading platform in Rust, smart contracts in Solidity. I keep measured results and estimates clearly apart.
$ pip install gpu-fit $ gpu-fit Qwen/Qwen3-8B --gpu L40S --context 32k --concurrency 4 Qwen/Qwen3-8B on 1× L40S (45.0 GiB) vLLM budget per GPU 41.4 GiB weights 15.3 GiB runtime overhead 1.9 GiB left for KV cache 24.3 GiB KV cache capacity 176,649 tokens Requests at full context 5.4 ✓ Fits. Suggested command: vllm serve Qwen/Qwen3-8B --max-model-len 32768
Where I work
Enterprise architecture and technology leadership in global pharma, where validation and change traceability are central.
AI platforms for transfer pricing and tax compliance: reading financial statements, benchmarking, drafting reports used for real filings.
Self-hosted LLMs with vLLM on H100, L40 and L40S: AWQ and FP8, shared-GPU budgeting, admission control and self-healing. Plus long-document AI and gateways with per-call cost tracking.
Market-making bots in Go, a trading platform in Rust, smart contracts in Solidity (EVM, Arbitrum) and Solana programs in Rust. Latency, keys, ledgers and edge cases first.
How I work
Every number I publish comes from a run on real hardware or arithmetic you can check. Estimates are labelled.
Money, ranges and percentiles are code with tests. AI does the reading and drafting, never the maths.
When data can’t leave, the model comes to the data: open models on your own GPUs, sized for the workload.
I lead teams and architecture, and I stay close to the code, the benchmarks and the logs.
Open source
How many users can this LLM serve on this GPU? A parameters-times-bytes estimate is a good start; gpu-fit goes further. It reads the model’s real architecture (grouped-query, hybrid, MLA, sliding-window, FP8) and mirrors vLLM’s memory budget. Then it hands you the command.
Also in the open: gpu-guard (Go, in development), a health check that flags a model whose speed drops, a common sign it has fallen back to the CPU.
Writing
Qwen3-4B and Qwen3-8B served with vLLM on two 24 GB GPUs from different generations. Why the 8B model won't start at 32K, why FP8 weights help more than a smaller context, and why both GPUs held the same amount of context.
Qwen3-32B on one NVIDIA L40S: why 16-bit weights don’t fit, why 8-bit weights still serve less than one user at 32K, and the two settings that change everything.
I’m always glad to compare notes on private LLMs, document AI, trading infrastructure or smart contracts.
Find me on GitHub