Code, design, infrastructure, and AI — from my first college projects to founding-engineer work.
2019 → today.
Runs the assistant end to end against simulated SaaS environments (email, accounting, e-commerce, team chat) generated from real tool schemas, with fault injection for realistic API failures. Scores tool-call correctness and final answers with deterministic checks plus an LLM-as-judge, and tracks regressions run over run against a full-tool-context baseline.
Every model call goes through one gateway backed by a central model registry, so the model for each task can be swapped without a deploy. Added end-to-end LLM tracing with PII redaction and per-tenant cost attribution.
Untrusted tool output is scanned before it enters the context window, kept separate from instructions, and truncated before it reaches downstream agents.
Database connection pooling, per-service connection limits, CPU-based horizontal autoscaling, and lease-based recovery for stalled background jobs.
Exactly-once metering of every model and tool call, usage-based pricing, a prepaid credit ledger with budget enforcement and auto top-up, and automated cost reconciliation. I built the billing service, the customer billing UI, and the billing console in Backoffice.
One place for the team to onboard MCP servers, monitor background jobs, manage feature flags, keep data clean, and track product usage.
Validates each new MCP server, runs onboarding autonomously, and publishes only the servers that pass health checks. Servers that support OAuth Dynamic Client Registration onboard with zero manual steps.
An OpenTelemetry pipeline that unifies logs, metrics, and traces across services — the foundation the LLM tracing was later built on.
Customers can export their own telemetry from Caylex to the observability tool they already use, configured from an admin UI.
Tracing on every model call: prompt, tokens, latency, cost. When an agent misbehaves, we can replay exactly what it saw. Built on the OpenTelemetry pipeline from the month before.
A new design system, built so AI agents can read it too (JSON-LD, llms.txt), with analytics. Live.
Agents pull in a skill — its instructions and tools — only when a task needs it, not on every request. The library can grow large while any single agent's context stays small.
Surfaces security issues found on connected servers and their tools, with a detail panel and advisory fixes, plus notifications when something needs attention.
The UI to give an agent its identity and instructions in plain markdown — its name, description, and how it should behave.
An admin sees how risky each tool is, sorts and filters by severity, bulk-disables the dangerous ones, and overrides a rating when needed — governance over what an agent is allowed to touch.
A capability map of what data lives in each connected tool (Linear, Notion, Slack), so the agent routes a query to the right tool instead of fanning out to all of them. I built the backend and the config UI. In September I rebuilt it with semantic search over the map, per-tool capability profiles (what each tool can and can't answer), and an eligibility classifier that decides which tools get mapped.
Works out what each server and its tools actually do, scores how much each tool can affect things, and flags security risks.
The flow for browsing the library of available servers, filtering by category, seeing details, and connecting them to a project — plus session recovery and faster page loads.
Agents that reason through multi-step problems and call the right tools, wired into scalable systems for industry-specific workflows.
An MCP server that pipes live logs and traces to the LLM — every debugging prompt arrives with real observability context, not just the question.
Improved root-cause-analysis speed by 60%, cutting incident resolution from hours to minutes for large-scale cloud environments.
Traces a failure back to its source and proposes the code or config fix — the triage pass an on-call engineer would otherwise do by hand.
Built the backend APIs and the UI for logs, the foundational pillar of an observability platform.
A Python framework that syncs third-party sources into BloomChat AI on a schedule, keeping its data fresh without manual imports.
Added Pydantic validation across the FastAPI endpoints, catching bad data and schema drift at the service boundary before it reached downstream pipelines.
Node.js, PostgreSQL, a vector DB, and AWS powering real-time image sharing for attendees at Laracon India 2023.
Kafka, FastAPI, and MongoDB on Azure, processing millions of live-stream events every hour.
Automated model training and data exchange across distributed nodes.
Deep-learning models that recognize furniture, trained and deployed on AWS.
Algorithms that rebuild full hyperspectral images from ordinary RGB input.
Integrated sensor data pipelines for live process monitoring on the factory floor, plus IoT-based permeability and viscosity tests using microcontrollers.
Automatic number-plate recognition in Python — detecting and reading vehicle plates from images with OpenCV.
Turning raw data into a clear narrative — a storytelling project (Alpha AI).
Before all this — college, 2017–18: a multi-purpose gadget, an audio mixer, and a CAD laptop stand.