// latest
View all projects ↓
Mixed-Traffic Vehicle Detection
100%rising_talent35+10+
// selected_work
Selected work
Mixed-Traffic Vehicle Detection
Read case study →
Golf Shot Detection Pipeline
Read case study →
Urdu News Video Analytics
Read case study →
PSL Sign Language Generation
Read case study →
Handwritten Urdu OCR
Read the story →
Diachronic Urdu Corpus
Multi-Source Video Retrieval
Pensioner Verification System
Sports Tracking Backend
Media Anonymizer
// demos
Systems running
Mixed-Traffic Risk Detection
Golf Shot Detection
Urdu News Analytics
PSL Sign Language
Handwritten Urdu OCR
// case_studies
How the flagships were built
case_01
Mixed-Traffic Vehicle Detection
Problem
COCO-trained detectors do not know what an auto-rickshaw is, and no public Pakistan-specific mixed traffic dataset existed to train one on. Roads here mix rickshaws, bikes, carts and pedestrians in ways Western datasets never see.
Approach
Recorded 13 hours of dashcam footage from my own vehicle and hand-annotated 293 frames in VoTT with a 4-class taxonomy, split by source video to avoid leakage. Face detectors failed on hazy frames, so I built an anonymization pipeline around person detection with a full manual plate audit, released as its own open-source tool. Fine-tuned YOLO26n on a laptop CPU in under 3 hours and wired it into a proximity risk overlay with tracking and debounced labels.
Pipeline
dashcam footage→VoTT annotation→anonymization→YOLO26n fine-tune→risk overlay
Result
203 anonymized clips, 1,506 boxes and the model, all public on Hugging Face. mAP50 0.72 on original frames, 0.63 on anonymized ones, both reported in the card.
case_02
Golf Shot Detection Pipeline
Problem
Golf coaches record full practice sessions, then scrub through hours of footage by hand to find each shot. Manual trimming does not scale past a few sessions.
Approach
A finite state machine watches the golfer through MediaPipe pose keypoints and flags candidate swings. YOLOv8 tracks the ball to confirm each candidate, and an audio impact detector validates the strike before a clip is cut.
Pipeline
raw video→pose FSM→ball tracking→audio validation→trimmed clips
Result
Fully automatic shot detection over long videos, delivered as a working pipeline. The project earned the Upwork Rising Talent badge.
case_03
Urdu News Video Analytical System
Problem
Pakistani news channels broadcast thousands of hours of Urdu content that nobody can search. Tickers, speech and speaker identity were all locked inside video.
Approach
Built every component from the data up: a self-annotated YOLO dataset for news ticker localization, a GAN-augmented Urdu OCR model, in-house ASR, speaker identification and summarization, then batch inference at scale at KICS-CLE.
Pipeline
broadcast footage→ticker detection→Urdu OCR→ASR + speaker ID→summarization→SQL database
Result
One full year of footage from the top 10 channels, processed end to end and queryable in SQL.
case_04
PSL Sign Language Video Generation
Problem
Deaf students in Pakistan have almost no signed video content. Producing it manually needs interpreters and a studio for every sentence.
Approach
Recorded a word-level Pakistani Sign Language dataset in collaboration with two deaf-education institutions. Trained a two-stage pipeline: diffusion-based gloss to pose, then cGAN pose to video. Now migrating to a StableAnimator diffusion pipeline with DWPose conditioning and SeedVR2 enhancement.
Pipeline
text gloss→pose diffusion→pose to video→SeedVR2 enhancement
Result
Paper under review at IEEE Access. Second-generation diffusion pipeline in progress.
// a personal one
Handwritten Urdu OCR
My father kept a handwritten Urdu diary from 1976 to 1985. Around 150 pages: poetry, mixed Urdu and Punjabi in Shahmukhi script, pages written in more than one direction, takhallus marks in the margins.
I am digitizing all of it. That meant designing a custom Label Studio annotation schema for poetry layouts and multi-orientation pages, then a TrOCR fine-tuning pipeline with a four-experiment ablation plan.
Target: an arXiv paper and a public inference API, so other families can read their pages too.
// research
Publications and research
- Scripting History: A Diachronic Urdu Text and Image Corpus from the 18th to 19th Centuries ↗published
- Pakistani Sign Language Video Generationunder review
- Handwritten Urdu OCR for Personal Archives
// services
Hire me for
Custom computer vision systems
Generative video and diffusion pipelines
End-to-end ML delivery
Technical mentoring and lectures
step_01
Send the footage
step_02
Feasibility demo
step_03
Fixed-scope build
// Proposals come with working demos, not slide decks. If the idea is feasible, you see it running first.
// client_feedback
What clients say
“He took my ideas and vision and turned them into a practical solution that made sense. The final outcome was perfect for an MVP and gives me a solid base to test with real users. His commitment to the project and genuine care about achieving the best outcome really set him apart.”
“Zeeshan delivered a thorough, well-documented final package for a genuinely hard problem (TTS for Tedim Chin, a low-resource tonal language). He was transparent about technical limitations rather than overpromising, communicated promptly, and the final handoff included everything needed to reproduce the work. Would hire again. 5/5.”
// about
Builder in Pakistan
I work from Lahore, Pakistan, across three roles: heading AI at Aswad Labs, running research at KICS-CLE at UET Lahore, and building my own company, Visaitech.
Most of my projects start with a dataset that did not exist. I collect the data, design the annotation, train the models and ship the API. The papers come from the same work that pays the invoices.
Kumrat Valley
At the lake, Kumrat
Northern Pakistan
Lecturing in Konstanz
// contact