Data engineering and production Claude applications
I build the analytics layer that holds an organisation’s operational data — and then put large language models on top of it. Independent practice, based in India, delivering to clients in the US and India.
What I do
Three service lines. Most engagements start with the first and grow into the others.
Turning operational exports into something you can actually ask questions of
Most organisations do not have a data problem, they have a data access problem — the numbers exist, spread across system exports nobody can join. I consolidate those into a queryable store, then build the reporting on top of it so the same number means the same thing everywhere.
Claude applications that stay affordable in production
Getting a demo working is the easy part. What decides whether an LLM feature survives contact with a budget is token discipline — caching, turn limits, model choice per task, and honest monthly accounting. That is the part I treat as an engineering problem rather than a line item you discover later.
Detection that earns its alerts
An alerting system people ignore is worse than none. Standard deviation bands fail badly on real operational data — it is volatile enough that a naive band either never fires or fires constantly. I build detectors on robust statistics, compare each entity against its own history rather than a fleet average, and tune against a backtest before anything goes live.
What would a Claude application cost you to run?
The question every AI project stalls on, and the one almost nobody will answer with a number. Pick a model and a volume to see it.
Monthly inference cost estimator Current Anthropic API rates, verified October 2026. Choose the shape of your workload and see what it costs.
From BI reporting to a production Claude application
The problem. Turnaround time was what the client sold. Work carried hard SLA commitments by priority, and the client had no reliable way to see breaches as they developed, compare performance across individuals and locations, or forecast the coming week’s volume. Reporting was manual and arrived too late to act on.
The outcome. A consolidated data layer feeding three automated reporting cadences, an anomaly detector that drafts notes for human review, and a Claude-backed interface that lets the client ask operational questions directly instead of waiting on a report.
The data layer. Raw operational exports consolidated into a queryable store feeding daily, weekly and monthly reporting — turnaround analysis, breach tracking against per-priority thresholds, per-entity leaderboards with week-on-week deltas, segment matrices, and next-week volume forecasting.
Why ±2σ had to go. Daily breach rates are volatile enough that a standard deviation band never trips. I replaced it with robust-z detection on median and median absolute deviation over a trailing window, plus a “worst day of the trailing period while materially beyond the median” rule. For some segments a long turnaround is chronic-normal, so the backlog detector compares each entity against its own trailing history rather than the wider population. Tuned against a backtest before going live.
The Claude layer. A natural-language interface over that data, deployed on Azure App Service. Prompt caching plus turn limits cut the running cost substantially, with nightly token accounting pulled from the API usage object into a per-month spend view — so the client sees what the feature costs before the invoice arrives.
Platform assessment. I assessed migrating onto Claude in Microsoft Foundry to consolidate model spend onto a single existing cloud invoice and keep prompts and completions inside the client’s own tenancy — a material compliance point when their customers ask where the data goes. That work covered the billing mechanics, the standard versus data-zone pricing trade-off, and the feature gaps that rule Foundry out for some workloads.
How I work
A one-person practice by design. You deal with the person doing the work.
How do engagements usually start?
Retainer or fixed-scope project?
What does the AI actually cost to run?
Does anything get sent automatically?
Will you name me as a client?
Do you work across time zones?
Working on something like this?
If you have operational data, no internal AI engineering function, and a question you cannot currently answer from it — that is the shape of problem I take on. Tell me what you are trying to see, and I will tell you honestly whether it is worth building.
rahul@vizzingdata.com