Senior software engineer with 15+ years in production banking, fintech, and telecom. Since 2024, focused on AI agents and evals: multi-agent systems, MCP, model routing, RAG, fine-tuning, LLM-as-judge, and guardrails.
Multi-agent systems with sharp tool boundaries, deliberate routing, and tested handoffs.
Picking the right model for the job, then squeezing cost and latency with fine-tuning and distillation.
Golden datasets, LLM-as-judge gates, and guardrails so agents stay reliable in production.
Cheap edge compute when it fits, real clusters when it does not.
Python and TypeScript drive most agent work today. Kotlin and PHP cover the production mobile and backend tail.
A snapshot of AI agents and platforms running today, drawn from Royan AB, SBAB Bank, and ParkUp Inc.
Golden datasets, regression back-testing, and pass/fail gates that catch agent-quality regressions before they ship.
Real-time, low-latency LLM agent loop with STT, MCP tools, TTS, and gesture sync. Output guardrails and prompt-injection checks on the deployed voice agent.
Multi-agent system with RAG over a vector database and embedding models, serving the parking inventory at scale.
LangChain pipeline that turns natural language into structured JSON, with schema validation and LLM-as-judge gating.
MCP servers, agent skills, and model platforms (including AWS Bedrock) integrated into the bank's engineering workflows.
Fine-tuned and distilled with Unsloth on Google Colab to cut inference cost and latency on the deployed voice agent.
The non-negotiables I bring to every project, from a one-off internal tool to a multi-agent system in front of paying users.
A slow, boring agent that always works beats a clever one that fails 10% of the time.
No eval, no merge. Golden sets and LLM-as-judge catch regressions before users do.
A fine-tuned 7B can beat a frontier model on cost, latency, and accuracy for narrow tasks.
Cheap models for easy turns, strong models for hard ones, judges only where they pay back.
Most agent failures are tool design failures. Tight schemas and small surfaces beat long prompts.
Input checks, output guardrails, and prompt-injection defense belong in the first commit, not the second incident.