Caching strategy
Route-level caching with TTL, versioned cache keys, invalidation on write events, stampede protection, and cache warming for high-traffic endpoints.
I build AI that ships — from notebook to production, for products people use every day.
Every project and post, embedded and mapped in 3D. Run a query to watch retrieval happen — the same idea that powers the RAG systems I build for teams.
const developer = {
name: "Deepak Kumar",
role: "Full-Stack Developer · AI/ML Engineer",
stack: ["Python", "PHP", "React/Next.js"],
focus: ["RAG systems", "Agentic AI", "Full-stack"],
coffee: Infinity,
openToWork: true,
};
// ideas in, products out
export default ship(developer);
Production fixes, automation scripts and side projects, most weeks of the year. Switch the editor to attention.py and click a word.
The flow most of my backend work follows: Client → API Gateway → stateless Laravel/FastAPI services → distributed cache → MySQL/Postgres (primary + read replicas) → Elasticsearch/pgvector → async queue workers.
Route-level caching with TTL, versioned cache keys, invalidation on write events, stampede protection, and cache warming for high-traffic endpoints.
Database create/update/delete events trigger async queue jobs that update Elasticsearch, keeping it eventually consistent with MySQL — with filtering, sorting and geo-distance queries, and a DB fallback if search is down.
Stateless services that scale horizontally, MySQL read replicas for high-read workloads, and Elasticsearch index lifecycle management for large datasets.
Logging, monitoring and alerting for queue lag, cache efficiency and search performance; zero-downtime deployments with rollback strategies and automated recovery.
LLMs integrated as just another backend dependency — structured JSON extraction, confidence-based escalation, and real-time streaming, wired into the same queues and caches as everything else.
Consolidated a split deployment onto one domain, cutting hosting cost roughly 50% — along with caching, indexing and query tuning that cut response times in several optimised flows.
Every AI suggestion sits behind a confirmation gate, enforced in code — a teacher finalises a grade, an annotator confirms a tag. High confidence never means automatic.
Not everything needs the cloud — Ermis V2 runs entirely on-device (local Whisper, SQLite, zero telemetry) for work where the data can never leave the machine.
Whether you’re hiring for your team or need a product shipped, these are the problems I solve end to end — from the first prototype to monitoring in production.
Assistants that answer from your own documents, agents with guardrails, and search that understands meaning.
CRMs, booking platforms and fleet/ops tools — built end to end on Laravel, React or Vue, with real cross-platform (Android/Electron) support.
The plumbing that keeps a team's pipelines honest: automated QA suites, serverless alerting and schedulers that replace fragile crontabs.
The product around the model: fast interfaces, solid APIs, auth and analytics — so users actually adopt it.
Local-first apps that use AI without sending data anywhere — on-device speech, vision or NLP, for work that can't touch the cloud.
ETL pipelines that turn messy, unstructured sources into validated, structured data at scale, with checkpoint recovery built in.
A multi-tenant Laravel ERP paired with a FastAPI AI service: AI question-paper generation from a leak-protected bank, and a rule-engine + LLM hybrid for answer evaluation — where the AI only ever suggests, never finalises a grade.
Real-time GPS tracking over WebSocket with OpenStreetMap routing and a dispatcher desktop companion.
38+ script automated QA suite across 4 regions, with an 8-trigger Azure Pipelines scheduler and serverless alerting.
Distributed ETL pipeline turning unstructured HTML into a validated 25+ field JSON schema, 100+ records/day.
Local desktop app: on-device Whisper drafting, a 26-rule compliance validator, and an append-only audit log. 100% local, no cloud.
Gemini AI drives the whole UI by natural language — compose, send, navigate, filter. Built in a 5-day hiring assessment.
Built from scratch, solo: real-time availability, payment-gated booking confirmation, full Laravel backend.
Multi-engine aggregator with Llama 3.3 content extraction, dual-layer caching, and real-time SSE streaming.
A simple loop that keeps ML projects honest: agree on what “good” means, get something working fast, measure it, then ship and keep watching.
Talk to users, map the data, and pick the one metric that proves the project worked.
output: success metric + planA working baseline in days, not months — real data, simple model, something people can click.
output: clickable demoTest sets, not vibes. Every change is scored for quality, latency and cost before it moves forward.
output: eval reportDeploy behind a flag, watch real usage, and keep improving with feedback from production.
output: live product + dashboardFine-tunes, notebooks, live demos and paper breakdowns — small experiments, kept in the open.
Open the full lab log"dependencies": {
"python": "daily",
"php": "daily",
"langchain": "production",
"hugging-face": "production",
"fastapi": "production",
"laravel": "production",
"pgvector": "production",
"next.js": "production",
"azure-devops": "production",
"go": "production",
"kubernetes": "learning"
}
The 37 tools I reach for across backend, AI/LLM, frontend and data — from a Laravel API to a RAG pipeline to the Postgres behind it. Pick a layer to light it up on the globe — or drag it to spin.
I enjoy the whole path from a notebook experiment to a reliable product — owning requirements, design, build, testing and deployment myself on most of the roles I've worked.
Right now I'm focused on production RAG systems, agentic workflows and full-stack delivery. Open to full-stack and AI/ML engineering roles, ideally on product-focused teams.
Download résuméOwn the public website (ASP.NET/C#/SQL Server/Umbraco) and built GScheck — a 38+ script automated QA suite across 4 regions, plus the Azure CI/CD behind it.
Sole engineer on a multi-tenant Laravel ERP + FastAPI AI service: RAG-grounded lesson planning, AI question-paper generation with a leak-protected per-school bank, and a rule-engine + LLM hybrid for answer evaluation. Also built the fleet/GPS tracking platform and a custom CRM.
Migrated search to Elasticsearch across two production sites (custom analyzers, synonyms, field boosts), with blue/green index rollover for zero downtime.
CQRS-style CRM: separated command/query paths, precomputed dashboard aggregates to a Redis-like cache, Elasticsearch fuzzy search on leads, and async CSV export via pre-signed S3 URLs.
E-commerce flows with cached catalog read models and a strictly ACID checkout path using the outbox pattern for notifications and invoice PDFs; seek-based pagination for large lists.
Hiring for an ML or full-stack role, exploring a collaboration, or just want to talk models? My inbox is open.