Ahmad ZushanAI · Systems · Platform

Available for AI platform, systems and security work

New in 2026: TabSpawn, Bol, Jugnu II and 27 more

00 · Ahmad Zushan · Delhi, IN · AI, Systems & Platform Engineer

AI systems that ship: the model, the platform and the machine underneath.

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.

Explore flagship work Technical dossier (PDF)
  • Vision & OCR
  • LLM agents
  • Model fine-tuning
  • Speech & language
  • eBPF & observability
  • Security engineering
  • Products, games & web
0documented projects
0VLMs fine-tuned on one 24 GB GPU
0requests per minute served
0builds shipped in 2026
  • Python
  • TypeScript
  • FastAPI
  • Next.js
  • Django
  • PyTorch
  • vLLM
  • PaddleOCR
  • Whisper
  • YOLO
  • Qwen-VL
  • Gemini
  • Bedrock
  • eBPF
  • Kubernetes
  • Redis
  • Kafka
  • PostgreSQL
  • nginx
  • Rust
  • WebAssembly
  • WebRTC
  • Three.js
  • FreeSWITCH
  • LoRA
  • OpenCV
  • systemd
  • Docker
  • Pillow
  • HarfBuzz
  • SQLite
  • Piper

01 · Flagship work

Systems designed, built and run in production.

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.

Architecture diagram of the document-intelligence platformProduction
№ 01 · AI · Vision & OCR

Haier Document-Intelligence Platform

One FastAPI platform, three use cases: handwritten-invoice extraction, long legal-PDF summarisation, and retail-display compliance scoring.

FastAPIPaddleOCRBedrockGeminiOllama
Live AI co-hostNode.js · Gemini Live API · InsightFace · Flask4production services on stage
Production · AI Conclave 2026
№ 11 · Products & Infra

QUIN — Live AI Event Co-Host

An AI that performs on stage with a human MC using real-time voice, plus a face-recognition reception desk: four production services.

Node.jsGemini Live APIInsightFaceFlasksystemd
Kernel-level data-flow visibilityeBPF · BCC/libbpf · ByteBuddy/ASM · Rust0.7%host CPU over millions of events
Case study
№ 14 · Security & Systems

eBPF Data-Flow Agent & Nucleus Platform

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.

eBPFBCC/libbpfByteBuddy/ASMRustKafka
Real-time telephony voice agentFreeSWITCH · mod_audio_fork · ESL · NLU2media paths: mod_audio_fork out, ESL back
Case study
№ 13 · LLM & Agents

Air India Voice Bot Platform

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.

FreeSWITCHmod_audio_forkESLNLUTTS

02 · The full index

Every project, one list.

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.

  1. 01

    Haier Document-Intelligence Platform

    One FastAPI platform, three use cases: handwritten-invoice extraction, long legal-PDF summarisation, and…    AI · Vision & OCRNov 2025 – Apr 2026→
  2. 02

    5,000-RPM Vision-LLM Serving

    Split the hottest compliance endpoints into a dedicated service sustaining 5,000 requests per minute against…    AI · Vision & OCRJan 2026→
  3. 03

    Retail-Compliance VLM Fine-Tuning Program

    99K-image dataset engineering and LoRA fine-tunes of six VLM families on a single A10G, to replace paid cloud…    AI · Vision & OCRMar 2026→
  4. 04

    Unified 3-Use-Case Compliance SaaS

    Turned three headless ML APIs into a polished Next.js + FastAPI product with auth, RBAC, jobs and hybrid…    LLM & AgentsApr + Jun 2026→
  5. 05

    Meter OCR — Gemini Reading Service

    Utility-meter reading API and front end: photo in, exact register reading and serial number out, with domain…    AI · Vision & OCRApr 2026→
  6. 06

    Meter OCR — Custom Detection + Recognition

    A zero-manual-labelling dataset via VLM × PaddleOCR cross-validation, then YOLO-OBB detection and fine-tuned…    AI · Vision & OCRApr – Jun 2026→
  7. 07

    PaddleOCR Microservice + Stamp Detection

    The ecosystem's OCR workhorse on two servers: orientation, structured boxes, markdown layout, parallel…    AI · Vision & OCRNov 2025 – Jun 2026→
  8. 08

    Document Preprocessing & OCR Toolkit

    Nine single-purpose tools that solved garbage-in for OCR: VLM sorter → ResNet-18 classifier → smart croppers…    AI · Vision & OCRNov 2025→
  9. 09

    Dynamic Dashboard — Visual Analytics over Salesforce

    Ask anything about CRM data in natural language; the LLM plans, fetches and composes the right dashboard…    LLM & AgentsApr 2026→
  10. 10

    Two-LLM Analytics Agent

    A stronger LLM writes guarded SOQL; a cheaper LLM shapes the results into components. The lean, cost-aware…    LLM & AgentsApr – Jun 2026→
  11. 11

    QUIN — Live AI Event Co-Host

    An AI that performs on stage with a human MC using real-time voice, plus a face-recognition reception desk:…    Products & InfraJun – Jul 2026→
  12. 12

    Infrastructure, DevOps & Disaster Recovery

    Two multi-service nginx estates, a systemd fleet, an instance-store data-loss incident and a recovery…    Security & SystemsNov 2025 – Jul 2026→
  13. 13

    Air India Voice Bot Platform

    A live airline voice bot: the caller speaks naturally, the system understands intent, calls real business…    LLM & AgentsReal-time telephony AI→
  14. 14

    eBPF Data-Flow Agent & Nucleus Platform

    Host→nucleus→backend platform that sees what apps send out over HTTPS, without decrypting TLS, at about 0.7%…    Security & Systems2025→
  15. 15

    LLM Security Layer — Prompt-Injection WAF

    A guardrail layer in front of LLM apps: an LLM-as-WAF classifier, tiny fine-tuned ONNX injection detectors…    Security & Systems2025→
  16. 16

    Crayto — Zero-Trust Browser Extension

    An MV3 anti-phishing agent that stops users typing corporate credentials into untrusted pages, enforced by…    Security & Systems2025→
  17. 17

    MMTC — Security & Infra Hardening at Scale

    Hardened about 48 VMs across a heterogeneous estate while keeping developer velocity: layered controls,…    Security & SystemsClient engagement→
  18. 18

    School Portal — Madrasa Management System

    A full-stack Django product: admissions, students, attendance, finance, reports and audit logging with…    Products & InfraProduction→
  19. 19

    Early NLP & Data Consulting (2024)

    Schema-aware NLP-to-SQL with GPT-3.5, Llama 2 + TF-IDF transcript classification, and retail forecasting: the…    LLM & Agents2024→

03 · Capabilities

What I bring across the stack.

Guiding principles: deterministic contracts around probabilistic models, cost-aware inference, resource-budgeted deployments, and observability from the kernel up.

01

Applied AI: vision, OCR & speech

OCR and vision-LLM extraction, retail-display scoring, classical CV and Whisper speech pipelines, fused with deterministic validators and context correctors.

PaddleOCRTrOCRYOLOQwen-VLWhisperOpenCV
02

Model training & serving

LoRA/4-bit fine-tuning on constrained GPUs, quantisation and high-throughput serving.

PyTorchPEFT/LoRAvLLMOllamaONNXbitsandbytes
03

LLM systems & agents

Planner/composer pipelines, structured outputs, tool and query guards, cost-tiered inference.

FastAPIstructured JSONRAGtraffic routingfailover
04

Security engineering

eBPF and runtime interception, LLM guardrails, Zero-Trust clients and fleet hardening.

eBPFByteBuddy/ASMprompt-injectionmTLSiptables
05

Cloud, MLOps & infra

Reverse proxies, process supervision, GPU hosts, streaming transport and data layers.

AWS/Azure/OCIKubernetesnginxsystemdKafkaRedis
06

Products, games & the web

Next.js/React and Django/FastAPI/Node products, static generators, WebAssembly and canvas games, WebRTC rooms, strict CSP deployments.

Next.jsTypeScriptDjangoPythonWebAssemblyWebRTC

04 · Journey

From NLP consulting to a product studio.

How I work

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 conversation
  1. 2024

    Applied NLP consulting

    Schema-aware NLP-to-SQL (GPT-3.5), Llama 2 + TF-IDF transcript classification and retail forecasting: the foundations of the later agent work.

  2. 2025

    Security & observability platforms

    eBPF outbound data-flow agent and nucleus, an LLM prompt-injection WAF, the Crayto Zero-Trust extension and MMTC fleet hardening.

  3. Nov 2025

    Document-AI platform in production

    Haier invoice, legal and compliance pipelines, the PaddleOCR microservice and a nine-tool preprocessing toolkit.

  4. Jan – Mar 2026

    Scale & self-hosted models

    5,000-RPM vision-LLM serving with failover, then a six-model VLM fine-tuning programme distilling cloud VLMs onto one A10G.

  5. Apr – Jul 2026

    Products, agents & live AI

    Unified compliance SaaS, Salesforce analytics agents, the meter-OCR pipeline and QUIN, a live AI event co-host with face-recognition reception.

  6. Aug – Sep 2026

    A self-hosted product studio

    FormReady, DocTools, ClipBridge, CutReady, RetroDeck, Emulator Hub, RetroDeck Live, Dusk Duel and a multi-tenant commerce platform, plus the msju.in and Healtharo client sites, all shipped and operated on self-managed servers.

  7. Sep – Oct 2026

    Games, voice and a bigger studio

    TabSpawn with seven original browser games, Jugnu II in 3D, Bol Urdu voice typing on the web and Android, Folio, MLTools, RetroDeck Live PS2, a 28-hour Urdu speech dataset, narrated product walkthroughs, and a server consolidation backed by a verified full backup.

05 · Documents

Case studies and the dossier.

Detailed write-ups with architecture diagrams, approaches tried and engineering decisions.

PDF

Haier AI Platform

Retail compliance, OCR extraction and legal pipelines: architecture and decisions.

Open PDF
PDF

Air India Voice Bot

FreeSWITCH + ESL real-time voice architecture, end to end.

Open PDF
PDF

eBPF Data-Flow Agent & Nucleus

Kernel-to-backend outbound visibility platform and Kubernetes rollout.

Open PDF
PDF

Technical Portfolio & Architecture Dossier

Cross-cutting summary of AI, security, cloud and MLOps work.

Open PDF
PDF

Crayto Development Roadmap

Zero-Trust browser-security agent: requirements and plan.

Open PDF

06 · 2026 builds · additions

30 things shipped since August.

Newer work from August to October 2026, listed separately from the portfolio above: products under zooshan.com, a games portal with original titles, Urdu voice and speech tooling, research notes, client websites and the server that runs them. Filter by kind; every entry opens a short case page, and live ones link out.

Products7Consumer tools built and operated under my own brand.

Live

DocTools — PDF Workshop

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.

FastAPIGhostscriptInkscapepypdf
doctools.zooshan.com№ 21 · 2026
Live

MLTools — 36 browser-only tools for ML and dev work

Annotation (boxes, polygons, OCR, NER, classification), dataset splitting and validation, fine-tuning calculators (VRAM, training time, learning-rate schedules), ONNX and tokenizer inspectors, and JSON, YAML and regex utilities, with nothing uploaded.

Next.jsTypeScriptTesseract.jsONNX
mltools.zooshan.com№ 24 · Sep 2026
Live

Zooshan.com — the tools hub

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.

HTML/CSSSEOnginxcertbot
zooshan.com№ 25 · 2026

Games & play8Original browser games, a games portal and the emulator rooms.

Voice & language5Urdu speech, reading and text tooling, from phones to training rigs.

Paused · server move

Bol — Urdu voice typing for the web and Android

A phone-first Urdu editor with document layouts and PDF export, Whisper transcription on the server or on the phone, and a context corrector that fixes sound-alike mistakes.

JavaScriptNode.jsWhisperwhisper.cpp
bol.zooshan.com№ 35 · Sep 2026 – present
Local app

Scribe Studio — local speech-to-text and OCR workbench

A tray-launched desktop app pairing faster-whisper with PaddleOCR: transcription, document OCR to Markdown, and a per-line Hindi + Urdu + English mode, all on a 4 GB laptop GPU.

FastAPIfaster-whisperPaddleOCRCUDA
no public site№ 37 · Sep 2026
Dataset

Urdu TTS dataset — 28 hours of aligned speech

A 30,128-clip, four-voice Urdu corpus built from public-domain sentences and synthesised through Azure Speech, packaged with checksums, review pages and a phase-by-phase roadmap for training a Piper or F5 voice.

PythonNumPySciPyAzure Speech
no public site№ 39 · Oct 2026

Tools, agents & research5Platforms, pipelines and measured experiments.

In development

QUIN Enterprise — voice agents that take action

The platform version of QUIN: vendors build voice agents that call APIs and databases through tools with confirmation policies, ground answers in their own documents, and design conversations in a visual flow studio.

TypeScriptFastifyPostgreSQL + pgvectorReact Flow
no public site№ 40 · Jul 2026 – present
Tooling

Narrated product walkthroughs — a video pipeline

Walkthrough and promo videos for ClipBridge, RetroDeck Live, FormReady, Bol and Jugnu II in English, Hindi and Urdu: scripted headless-Chrome recordings, AI narration, captions, and a frame-by-frame check against the words.

Node.jsChrome DevToolsGemini TTSffmpeg
no public site№ 41 · Oct 2026
Research

pagemap — a compact page map for a self-trained browsing agent

Step one of a browsing agent that runs on my own model: code, not an LLM, turns any rendered page into a short text map of roles, names and targets that a small fine-tuned model can act on.

PythonPlaywrightJavaScriptfine-tuning
no public site№ 42 · Sep 2026
Research

CPU vs GPU asymmetry — a measured study

Can four laptop cores outperform a GPU? Only through asymmetry: an inverted-file retrieval kernel beat the GPU's roofline floor 18.9× at 100% recall, and the write-up says exactly where the trick stops.

CUDANumPybenchmarkingctypes
no public site№ 43 · Sep 2026
Lab note

ESP32 aurora — a MicroPython display experiment

An ESP32 driving a 128×64 OLED over I²C with a drifting aurora animation in MicroPython: the smallest project here, kept because the fix was the oldest kind.

ESP32MicroPythonSSD1306I²C
no public site№ 44 · 2026

Client sites3Websites designed and shipped for clients in 2026.

Live

Siraj-ul-Uloom — a trilingual heritage website

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.

PythonHTML/CSSRTL typographyPillow
msju.in№ 45 · Sep 2026
Staging

asadisraili.com v2 — a marble-and-inlay redesign on staging

A new identity for a scholar's lecture website, marble and walnut with turquoise and carnelian touches and a pointed-arch motif, built as a separate static generator and shown on a staging host before anything in production changes.

PythonHTML/CSSRTL typographynginx
staging.asadisraili.com№ 47 · Sep 2026

Platform & operations2The hosting layer underneath them.

Production

One server, twenty-two sites — the Zooshan hosting platform

Every product, game and client site here runs on a single 2-vCPU virtual machine: twenty-two nginx virtual hosts, Docker Compose and systemd services, Let's Encrypt certificates, release directories and tested rollback.

nginxDocker Composesystemdcertbot
no public site№ 48 · Aug – Oct 2026
Operator-only

Operations dashboard — one page for every site on the server

A read-only page listing every site and application on the server with bot-filtered traffic, the release that is live, certificate expiry, systemd services, containers and resources, regenerated hourly behind basic auth.

PythonnginxsystemdSVG
no public site№ 49 · Sep 2026

07 · Contact

Let's build something that ships.

Open to AI/ML platform roles, systems and security engineering, and focused product work. Based in Delhi, working across remote GPU and cloud estates.

Phone / WhatsApp
+91 97200 40070
Location
Delhi, India · IST (UTC+5:30)