Work
Every engagement below ended with a measurable before-and-after. Challenge, approach, result — no anonymous logos, no vague "improved efficiency".
Portfolio
↓ 95% dashboard load time
20-second dashboards became sub-second — and beat PowerBI head-to-head.
A production analytics platform for an enterprise energy client: rebuilt query layer, two-tier Redis cache at a 94% hit rate, and an ECharts Canvas UI rendering 50,000+ points at 60fps. Selected over PowerBI in the client's direct evaluation.
42,000ms → 380ms
A 110x faster critical query at one billion rows.
Benchmarked ClickHouse against PostgreSQL and TimescaleDB on the real 1B-row dataset, designed MergeTree schemas and materialised views around the actual query patterns, and migrated the critical path. The full decision process is public.
Read the write-up< 200ms voice latency
Real-time AI voice agents in production.
A production voice-AI pipeline with sub-200ms end-to-end latency: real-time WebRTC transport via LiveKit, streaming speech handling, and LLM-driven conversation — the response starts before the caller finishes expecting one.
Live production SaaS
A market-analytics platform, running publicly right now.
A full SaaS product — data ingestion, analytics, and a real-time dashboard — deployed on Vercel and Railway with the source open for inspection. Not a demo: real users, real uptime.
View project↑ 25% MAU in 3 months
Data-driven product features that moved retention.
A consumer relocation platform where data-informed feature work — smarter recommendations, faster flows, measurable iteration — grew monthly active users by a quarter within one quarter.
↓ 40% QA debugging time
AI-assisted analysis where the hours actually go.
Applied-AI tooling that cut quality-assurance debugging time by 40% — automated triage and analysis of failures so engineers spend their time fixing, not hunting.
↑ 30% operational efficiency
Workflow automation with a measured payoff.
Operational workflow automation for a client team — structured pipelines and intelligent routing that lifted overall operating efficiency by 30%.
Public benchmarks
Architecture decisions for IoT-scale data, in the open.
The reasoning, trade-offs, and benchmark numbers behind IoT-scale data architecture choices — published so you can inspect how we make decisions before you hire us to make yours.
View projectOur expertise
Model selection, fine-tuning where it earns its keep, and evaluation harnesses — production AI, not demos.
Tool-using agent workflows and sub-200ms real-time voice pipelines, shipped to production.
Retrieval pipelines with tuned vector search, grounded citations, and measured precision/recall.
Forecasting, anomaly detection, and decision-support on billion-row datasets.
ClickHouse, Spark, and Azure Data Lake pipelines that stay fast as data grows.
Azure (AKS to AD), AWS, Vercel — CI/CD with staged rollout and automated rollback.
42,000ms → 380ms
20s → <1s
50,000+ @ 60fps
17 engineers · 95%
In the open