Applied AI & vision
OCR and vision-LLM extraction, retail-display scoring and classical CV, fused with deterministic validators.
Available for AI platform, systems and security work
00 · Ahmad Zushan · Delhi, IN · AI, Systems & Platform Engineer
AI/ML engineer working end to end: OCR and vision-LLM pipelines, fine-tuned models, LLM agents behind guarded APIs, kernel-level observability, LLM security, and the nginx, systemd and GPU infrastructure it all runs on. I ship the model and the platform around it.
01 · Flagship work
Document AI at scale, a live AI stage co-host, kernel-level data-flow security and real-time telephony. Case-study PDFs are attached where they exist.

One FastAPI platform, three use cases: handwritten-invoice extraction, long legal-PDF summarisation, and retail-display compliance scoring.
An AI that performs on stage with a human MC using real-time voice, plus a face-recognition reception desk: four production services.
Host→nucleus→backend platform that sees what apps send out over HTTPS, without decrypting TLS, at about 0.7% host CPU over millions of events.
A live airline voice bot: the caller speaks naturally, the system understands intent, calls real business APIs, and speaks the answer back into the same call.
02 · The full index
Nineteen documented pieces of work from 2024 to 2026, in order. Filter by discipline; hover a row for a preview; open any row for architecture, hard problems and what shipped.
03 · Capabilities
Guiding principles: deterministic contracts around probabilistic models, cost-aware inference, resource-budgeted deployments, and observability from the kernel up.
OCR and vision-LLM extraction, retail-display scoring and classical CV, fused with deterministic validators.
LoRA/4-bit fine-tuning on constrained GPUs, quantisation and high-throughput serving.
Planner/composer pipelines, structured outputs, tool and query guards, cost-tiered inference.
eBPF and runtime interception, LLM guardrails, Zero-Trust clients and fleet hardening.
Reverse proxies, process supervision, GPU hosts, streaming transport and data layers.
Next.js/React and Django/FastAPI/Node products, static generators, WebAssembly in the browser, strict CSP deployments.
04 · Journey
Spec first, mock-first processors, then real inference behind an unchanged contract. Everything I build I also deploy, monitor, back up and, when it breaks, recover.
Start a conversationSchema-aware NLP-to-SQL (GPT-3.5), Llama 2 + TF-IDF transcript classification and retail forecasting: the foundations of the later agent work.
eBPF outbound data-flow agent and nucleus, an LLM prompt-injection WAF, the Crayto Zero-Trust extension and MMTC fleet hardening.
Haier invoice, legal and compliance pipelines, the PaddleOCR microservice and a nine-tool preprocessing toolkit.
5,000-RPM vision-LLM serving with failover, then a six-model VLM fine-tuning programme distilling cloud VLMs onto one A10G.
Unified compliance SaaS, Salesforce analytics agents, the meter-OCR pipeline and QUIN, a live AI event co-host with face-recognition reception.
FormReady, DocTools, ClipBridge, CutReady, RetroDeck and Emulator Hub, plus the msju.in and Healtharo client sites, all shipped and operated on a single self-managed server.
05 · Documents
Detailed write-ups with architecture diagrams, approaches tried and engineering decisions.
Retail compliance, OCR extraction and legal pipelines: architecture and decisions.
Open PDFFreeSWITCH + ESL real-time voice architecture, end to end.
Open PDFKernel-to-backend outbound visibility platform and Kubernetes rollout.
Open PDFCross-cutting summary of AI, security, cloud and MLOps work.
Open PDFZero-Trust browser-security agent: requirements and plan.
Open PDF06 · Live sites · additions
Newer work from August and September 2026, listed separately from the portfolio above: products under zooshan.com, two client websites and the single VM that runs them. Each entry opens a short case page and links to the live site.
Consumer tools built and operated under my own brand.
Tell it what the portal asks for (JPG, 200×230 px, 20–50 KB) and get a file that passes: photos, signatures, scans, PDFs and the numbers portals ask for.
Thirteen server-side PDF and vector tools in one focused workspace: merge, split, compress, protect, unlock, watermark, number, clean metadata, convert, and trace images into real vector paths.
Private text, links and files shared between your own devices, with connections between accounts, directed transfers, live updates and single-use guest links.
Calligraphy, stickers, vinyl, laser and CNC tools that turn a typed name into a file the machine accepts, entirely in the browser.
The front door for the product family: it explains which tool fits which job, hosts long-form guides on application photos, PDFs and file sharing, and carries the studio's privacy commitments.
Load your own PlayStation disc images and BIOS and play privately in a modern browser; nothing is uploaded.
One page for many systems: arcade, Nintendo, Sega, Sony, Atari and classic computers, each running lawful backups locally in the browser.
Websites designed and shipped for clients in 2026.
A redesigned website for a 124-year-old madrasa in Sambhal: heritage photography, a 52-image gallery, admissions and donation guidance, and full Urdu and Arabic editions.
A 67-page catalogue site for a Sambhal herbal manufacturer: product families, compositions, classic Unani ranges, quality pages and WhatsApp-first contact.
The hosting layer underneath all of them.
Every product above runs on a single 2-vCPU virtual machine: twelve nginx virtual hosts, Docker Compose and systemd services, Let's Encrypt certificates, release directories and tested rollback.
07 · Contact
Open to AI/ML platform roles, systems and security engineering, and focused product work. Based in Delhi, working across remote GPU and cloud estates.
ahzushan@gmail.com