AI Integration
OpenAI, Claude or Gemini wired into products that already exist — or open-weight models self-hosted via Ollama when data can't leave your network.
visitor@dk:~$ whoami
role=Full-Stack Developer & Security Specialist
visitor@dk:~$ cat summary.txt
I'm a full-stack software engineer with 13+ years building production web applications end to end — architecture, APIs, interfaces and deployment. I ship software that's secure by design, with AI integrated where it genuinely moves the needle — from lean small-business sites to enterprise platforms built to scale.
available for new projects
Three disciplines, one habit of mind — build it well, add AI where it earns its place, then assume someone is already trying to break it.
Thirteen years in, I've built at most of the sizes the web comes in — a single-page site for someone who just needed to be findable, brochure sites and online stores for small businesses, internal applications for teams that had outgrown their spreadsheets, and product platforms where an hour of downtime is a number somebody reports upwards. The work changes shape at every one of those scales. What doesn't is working out what the business actually needs before writing a line of it.
What I add on top is a security habit most developers pick up late. I've cleaned up sites after they were compromised and hardened the servers so it didn't happen twice, and spent part of my career building software for network security products — which is where I learned what an attack looks like from the defender's side.
More recently that's grown to include AI features inside products that already exist — hosted LLMs where that's the right call, self-hosted open-weight models when data can't leave the network — and CI/CD pipelines that take the drama out of releasing. I work best with founders and small teams, and I'm as comfortable taking over a codebase someone else started as writing one from scratch.
OpenAI, Claude or Gemini wired into products that already exist — or open-weight models self-hosted via Ollama when data can't leave your network.
End to end — architecture, database, API, interface, deployment. PHP and Laravel through to Angular and Node.
Threat modelling, secure code review and hardening baked into the build, not bolted on before launch.
GitHub Actions pipelines that build, test and deploy on merge — so releases stop being an event.
Grouped by where each one lives in the stack, from AI down to the network.
Eight things people hire me for — from adding AI to a product to cleaning up after someone broke into one.
Real AI capability added to products that already exist. OpenAI, Anthropic Claude or Google Gemini where a hosted API is the right call — or open-weight models running on your own server when customer data can't leave the network. Either way, automation that removes manual steps rather than a chat box bolted on the corner.
Angular, React, Node.js and PHP applications built end to end — architecture, database, API, interface, deployment. Small business sites through to enterprise systems.
GitHub Actions pipelines that build, test and deploy on merge — so releases stop being an event and start being a Tuesday.
Applications built to hold up in the open: input validation, authentication and access control done properly, and a security review before launch rather than after the incident.
Cleaning compromised sites, closing the way back in, and leaving the server configured so it doesn't happen a second time.
Security audits, malware cleanup and hardening for the platforms attackers target most — vulnerable plugins and extensions patched, admin access locked down, and bugs fixed without breaking the store.
Stores built, customised and fixed on Shopify, Magento, BigCommerce and OpenCart — themes, extensions, payment and shipping integrations, and the checkout bugs that quietly cost sales.
Interfaces for firewall and network security tools — dense configuration screens and alert data made legible to the people on shift.
Open-source tools I've built and shipped, plus client work. Problem, approach, and what it actually does.
Knowing what is actually on a network usually means putting an agent on every host or running active scans that set off alarms. Both are intrusive, and both are a project in their own right before you learn anything.
A Go tool that reads Suricata's eve.json and correlates protocol metadata — DHCP, ARP, TLS, Kerberos, SMB, SNMP, SSH — into device profiles carrying a confidence score and an evidence trail. No agents, no scanning, no ELK stack to stand up.
Builds a passive asset inventory, maps users to devices, validates network segmentation policy and flags changes as they happen. Produces self-contained HTML reports and runs on Linux, FreeBSD (OPNsense and pfSense), macOS and Windows.
Most Microsoft 365 security tools hand you a point-in-time score. They can't tell you whether your posture is improving or quietly sliding — and Microsoft only keeps Secure Score history for about 90 days.
A multi-tenant Laravel and Vue dashboard that polls Microsoft Graph with application-only authentication, running independent collectors on staggered schedules and taking daily snapshots so history outlives Microsoft's retention window.
Tracks MFA registered versus actually enforced with separate admin coverage, active Conditional Access policies, and device compliance across Entra ID and Intune — then alerts on configuration drift when a control moves the wrong way. Role-based access, branded PDF reports and scheduled email.
Software vendors want an AI assistant sitting on their documentation, but sending customer questions to a third-party API is a non-starter for many of them — on cost, on privacy, or on both.
A retrieval-augmented chatbot that runs entirely on the vendor's own infrastructure: local models through Ollama or vLLM, embeddings via fastembed, vectors in sqlite-vec and sessions in Redis. No API keys, and no data leaving the network.
Ingests Markdown, HTML, PDF and plain text, then answers with streamed responses and source citations. When retrieval confidence is too low it declines to answer rather than inventing something. Ships as an embeddable widget with custom branding, rate limiting and GPU concurrency caps.
Something like these on your list? Let's talk about it
I've been thanked and rewarded by these companies for responsibly reporting security vulnerabilities in their platforms.
Security Researcher Acknowledgement
View acknowledgementSecurity Researcher Acknowledgement
View acknowledgementSecurity Researcher Acknowledgement
View acknowledgementSecurity Researcher Acknowledgement
View acknowledgementThe things people most often want to know before getting in touch.
Deepanshu Kapoor is a freelance AI, full-stack and security engineer with 13+ years of experience. He builds and secures web applications, integrates AI and LLMs (including self-hosted models and RAG), removes website malware, and has built frontends for firewall and IDS/IPS security products. He is acknowledged in the security researcher hall of fame by Microsoft, eBay, Red Hat and Sellfy.
I add AI to existing products using hosted APIs (OpenAI, Anthropic Claude, Google Gemini) or self-hosted open-weight models via Ollama and vLLM when data cannot leave the network. This includes retrieval-augmented generation (RAG) over private documents, embeddings, and FastAPI backends — focused on automation that removes manual steps rather than a bolted-on chat box.
Yes. I clean compromised WordPress, Magento, OpenCart and custom sites, identify and close the entry point, and harden the server so the site is not reinfected. I also perform malware analysis and full security audits.
Frontend: React, Angular, Next.js, JavaScript. Backend: PHP, Laravel, CodeIgniter, Node.js, Python, FastAPI. AI: OpenAI, Claude, Gemini, Ollama, RAG, embeddings. Databases: MySQL, PostgreSQL, MongoDB, GraphQL. Ecommerce: WordPress, Magento, Shopify, BigCommerce, OpenCart. DevOps & security: AWS, Docker, Linux, Nginx, GitHub Actions, OWASP Top 10, Suricata IDS/IPS.
Yes. I'm open to freelance and full-time work for clients worldwide, work directly with clients rather than through a team, and typically reply within an hour. You can reach me by email, LinkedIn or WhatsApp.
visitor@dk:~$ ./start_project.sh
Tell me what you're building — or what's broken. I'll come back within an hour with an honest answer about whether I'm the right person for it.
Building something new, or worried about something you've already built? Either way, I read every message.