I engineer software and AI agents used by millions.
I care about the parts of engineering that outlast a launch: architecture that holds up at scale, the right database for the job, async pipelines that don't fall over, and code another engineer can pick up six months later.
Most of my recent work is agentic AI: orchestrators and subagents, tool calling, structured outputs, evals and the telemetry that keeps them honest in production.
Actual production figures as of October 2026
PosterMyWall is an online design and marketing platform. I've spent 3+ years on its editor and AI products: the multi-agent AI Chat, the AI Websites beta and an R&D poster pipeline, plus the media pipelines and canvas tools underneath them.
Usage of AI features I engineered
Actual production figures as of October 2026
Live since Feb 2026. Core engineer.
An agentic marketing suite that takes a business from idea to published campaign in one conversation. An orchestrator agent routes each turn to specialist subagents with 40+ tools on an extended NeuronAI framework, and streams back interactive UI such as surveys, option pickers and draft cards right inside the chat.
Actual production figures as of October 2026
| Month | Positive | Ratings |
|---|
Beta. Lead engineer and architect.
An agentic website builder inside AI Chat that turns a conversation into a complete, editable website, streamed onto the canvas as the model writes it.
R&D, not launched. Architect.
Drag the divider to compare. Pick a template to see its brief. Prompts are illustrative.
An agentic pipeline that turns a one-line brief into finished posters by re-theming designs from PosterMyWall's library of 4M+ templates, without breaking the designer's layout.
The pipelines and editor tools underneath the AI products.
Image uploads took a median 22.5s. I traced about 90% of the wait to fixed 2-second polling across ten sequential renderer calls, collapsed them into one queued job with sub-second polling, and shipped a cache-key fix through a zero-downtime, three-step rollout.
For video, I instrumented each step in CloudWatch first, then replaced a five-step renderer chain with one job that encodes the preview once at 850px and rebuilds high-res files from the original, which also made exports sharper. The p90 for video fell from 97s to 37s.
Subtitling for audio and video, which I architected: AI transcription with OpenAI, a React timeline for syncing and editing each line, and a custom Fabric.js grouped object that renders styled, animated, word-highlighted captions without slowing the canvas.
A text-to-image panel with prompt and style controls, backed by several image models with prompt validation, per-user quotas and Stripe billing. Results drop straight onto the design canvas.
I scaled stock media across 8 categories by reworking the MySQL schema, automating preset generation on S3 with ImageMagick and FFmpeg, and adding AI keyword and category tagging for search. A Python computer-vision tool cuts individual stickers out of large sheets, then upscales them and removes backgrounds, taking a manual process of about 50 stickers a day to 100+ in minutes. An admin dashboard generates AI stickers at scale with context-aware keywords.
I reverse-engineered the 2D transformation math behind Fabric.js's resize, scale and rotate controls and rebuilt them as HTML/CSS overlays that work even on objects outside the visible viewport. I also built a constraint-aware colour algorithm that applies and reshuffles palettes across layered designs in real time, using z-order and overlap analysis to keep touching items in contrast, and refactored 57 React components into reusable hooks and selectors to stop wasted re-renders.
SuperOrgs is the system of record for the hybrid workforce: AI agents on the org chart next to people, each with an owner, a cost and a lifecycle. I engineer the core platform, now live with real customers, and have shipped 108 PRs across 74 tickets in about 3.5 months, owning features from design doc to production.
Tool-calling LLM agents mine six work tools, including Linear, Slack and Asana, for automation opportunities under strict consent, discard-after-use and per-org spend caps. Ideas are scored 0–100, deduplicated with pgvector and turned into documented processes from uploaded recordings (Whisper and Claude) that experts sign off on without logging in.
Every agent run is priced against a seven-part loaded human-cost model on a nightly ledger, surfacing time saved and value per run ("6 min saved, worth $5") and sending monthly PDF reports by email and Slack. A real-Postgres test suite reproduces a worked example to the cent.
Rebuilt a page that froze for 10+ minutes around a bounded background job. Got a 118K-conversation Slack workspace syncing by cutting about 590 API calls to 120, so 4,010 channels sync in 3.4 minutes instead of about 30. Took a hot query from 2.9 ms to 0.03 ms with targeted indexing.
Cross-run agent handoff chains, AI prioritization that never overrides a person's call, a screen-recording intake pipeline that turns a recorded task into a step-by-step workflow with screenshots, and a one-click Claude Code brief for building new agents.
Freelance and personal builds, and research from my time at LUMS. Open a row for the details.
Independent
Real-time air-quality readings, 24-hour ML forecasts and personalised, location-based push notifications. I led UI/UX and full-stack. When monitor readings went minute-level, I migrated them from MongoDB to InfluxDB, cutting storage costs by 30× and making time-range queries 200× faster.
Freelance. A fully responsive marketing site with scroll reveals, a typewriter hero, a testimonial ticker and case-study pages.
Personal. Level filters, a daily revision mix (focus topic, phrase of the day, flip-card word sprint, writing prompt) and an 11-section worksheet with answers you can reveal. Built with Claude as a single-page app.
It pulls the transcript and answers with OpenAI, and messaging between content and background scripts removes two steps that similar tools require.
Research at LUMS
Final year project supervised by Prof. Ihsan Ayyub Qazi, an HCI study of whether fact-checking portals counter political misinformation. A Python scraper pulled and Urdu-translated articles from Soch Fact Check, with cron-scheduled translation and Redis caching that cut API costs by 95%. The React and Flask web app was tuned for low-end phones, tracked user journeys and went out weekly to 2,000+ participants over WhatsApp.
ML research supervised by Prof. Ihsan and Zafar Ayyub Qazi in the Network and Systems Group. I explored tokenization schemes for feeding C++ into a transformer, built a corpus of AI solutions to compare against a human user study, and automated code metrics such as Halstead volume, cyclomatic complexity and maintainability index. A Mann-Whitney U test found no statistically significant separation between the two.
Security research that bypassed Android 12's 200 Hz motion-sensor sampling limit through sensor-management vulnerabilities. A kernel-level interleaving algorithm doubled the sampling rate without user permissions, recording accelerometer data alongside Google Assistant audio to produce clean mel-spectrograms.
HCI research at the CHISEL lab, supervised by Prof. Suleman Shahid: a gamified app teaching sign language to hearing primary-school children. I ran interviews with educators and parents, contextual inquiries with 32 children and storyboarding sessions, then built and usability-tested a high-fidelity prototype: 15 of 18 children completed every task.
Research at LUMS' Center for Speech and Language Technologies, supervised by Prof. Agha Ali Raza. I compiled a phonetically diverse Urdu corpus, coordinated a 13-person team to gather 18,000+ recordings, fine-tuned Wav2Vec2 with CTC to a 37% word error rate, and helped bootstrap a standard Urdu NLP library for the lab.
This section is maintained by THSort-Agent, an autonomous agent I built and operate. It works under its own GitHub identity and opens its own pull requests, and every item below links to verifiable proof.
No merged work yet — the agent is on its first run. Check back soon.
I'm open to full-time roles, and I take on a few contract projects at a time: AI product development, React and React Native builds, Node and Python backends, and performance work on existing codebases. I usually reply within a day.
