Case studies

Work I've led, and what it delivered.

These are programs I led as an engineering leader. The same playbooks now power my client engagements.

Top-10 US bank (Capital One)

30,000+

users on Claude Code, Codex, Cowork & Windsurf

AI coding agentsGovernanceAWS Bedrock

AI coding agents for 30,000+ engineers, inside a regulated bank

Challenge

Engineers wanted the latest AI coding agents; Risk and Cyber needed proof that source code and sensitive data would stay protected, that usage was auditable, and that costs wouldn't run away.

Approach

Routed every tool through an enterprise gateway on AWS Bedrock with OAuth, RBAC and per-user, per-minute token limits.
Shipped centrally managed settings, project memory standards, hooks and MCP configuration so guardrails travelled with the tool.
Instrumented usage with OpenTelemetry for cyber logging and developer-productivity metrics.
Built a fast, tested release cadence so new models reached engineers in days, not quarters.

Outcome

30,000+ users across Claude Code and Cowork, with Codex and Windsurf alongside
Cost held down with prompt caching, model routing and effort-level tuning
A pattern Risk, Cyber and the Model Risk Office could approve and reuse

Top-10 US bank (Capital One)

< $4

median cost per reviewed PR at 350K+ PRs / month

Agentic AIEvalsDeveloper platforms

An agentic code reviewer across 50,000+ repositories

Challenge

Generic AI review bots are noisy and expensive. Engineers ignore them after a week. The bank needed findings aligned to its own standards, at monorepo scale, in every language.

Approach

Built on the Claude SDK with internally curated standards so every finding maps to a rule.
Used deterministic checks around the model to suppress hallucinated findings.
Made harness, model, rules and agents swappable, with a self-service rules-testing pipeline.
Gated changes on offline evals: accuracy, false-positive rate and dissatisfaction rate.

Outcome

50,000+ repositories and 350K+ PRs per month
~10-minute SLA to first review on serverless infrastructure
Headless agents extended to tests, docs and gap analysis

Top-10 US bank (Capital One)

80%

of repeated support queries resolved

RAGAgentsKnowledge platforms

DeX: an agentic RAG assistant that knows when not to answer

Challenge

Platform teams were drowning in repeat questions across Slack. Answers existed — scattered across docs and years of threads — but a naive chatbot would confidently invent the rest.

Approach

Ingested product docs and historical Slack conversations with bge-m3 embeddings and a reranker.
Added answerability, query-rewrite, critique and summary steps so the agent declines when the source can't support an answer.
Rolled out in shadow mode first, then promoted confident answers to the main channel.
Tracked coverage, accuracy, freshness, MTTA and MTTR on OpenTelemetry dashboards.

Outcome

Scaled to 70+ platforms as a multi-tenant, self-serve service
Automated documentation generated from resolved Slack threads
A human hand-off path whenever confidence was low

Paytm

$1.8M+

monthly cloud savings

FinOpsPlatform engineeringData

$1.8M a month out of a 55PB data platform

Challenge

A fast-growing fintech data estate — 15PB warehouse, ~10,000 analysts — was running always-on clusters and paying for cross-region traffic it didn't need.

Approach

Migrated EMR fleets to AWS Graviton at roughly a third of the price.
Replaced always-on Hive/Hadoop with autoscaling Starburst on Kubernetes.
Moved to transient, event-driven EMR so not even master nodes ran 24/7.
Removed inter-region data transfer by replicating clusters per region.

Outcome

$1.8M+ saved every month
Provisioning cut from days to ~1 hour; access from days to ~10 minutes
The same playbook now applies to LLM token economics

Work with me

Want results like these in your organisation?

A free 30-minute call to understand your context. If I'm not the right fit, I'll tell you — and point you to someone who is.

Book a call er.kushbatra@gmail.com WhatsApp LinkedIn