About
Hi, I’m Kiran.
I build systems where accuracy matters: AI for regulated industries (life sciences, tax & finance) and digital-asset trading infrastructure. What connects them is that an error, or a slow answer, has a real cost, so the systems need to be accurate, explainable, auditable and fast.
What I do
- Life sciences. Enterprise architecture and technology leadership in global pharma, where validation and change traceability are central.
- Tax & finance. AI platforms for transfer pricing and tax compliance: reading financial statements, benchmarking and drafting reports that go into real filings.
- AI in production. Self-hosted LLMs with vLLM on H100, L40 and L40S (AWQ and FP8, shared-GPU budgeting, admission control, self-healing when a GPU drops out), document AI for 100+ page reports, and gateways that route between cloud and private models with per-call cost tracking.
- Digital assets. Market-making bots for exchanges in Go, a trading platform in Rust, smart contracts in Solidity (EVM, Arbitrum) and on-chain programs in Rust (Solana).
- Cloud. AWS platforms end to end: networking, security, GPU fleets and CI/CD.
I lead teams and architecture, and I stay close to the code, the benchmarks and the logs.
Languages and tools: Go · Rust · Solidity · TypeScript · Python · PostgreSQL · vLLM · AWS
What I believe
AI should do the reading and the drafting. Money, ranges and percentiles should be code with tests.
If the data can’t leave, bring the model to the data, and size the hardware with arithmetic anyone can check.
What I write about here
- GPU sizing: what fits on which GPU, at what context length, for how many users.
- vLLM in production: the settings that matter and the failures worth knowing about.
- Cost: self-hosting against APIs, per million tokens, and what quantisation really saves.
- Regulated AI: audit trails, determinism and keeping humans in the loop.
- Trading infrastructure: latency, rate limits, risk controls and smart-contract safety.
Numbers here come from my own runs on public models or from arithmetic shown in full. When something is an estimate, it says so. Views are my own and don’t represent any employer or client.
Elsewhere
GitHub · gpu-fit on PyPI · gpu-guard · Medium · RSS