Habeeb
Computer Vision
Detection, segmentation, pose estimation and 3D perception — plus the unglamorous half: latency budgets, quantisation and models that actually survive a real input.
Computer Vision Engineer
Vision systems that see.
Models that reason.
habeeb 0.99
I'm Habeeb — a computer vision engineer who ships vision systems into production: medical imaging in hospitals, perception on autonomous vehicles. A lot of that work happens at the edge, squeezing transformer detectors and 3D perception onto hardware that isn't a datacenter. These days my curiosity goes into vision–language models and post-training.
- Projects
- 4
- Essays
- 23
- Articles
- 9
01 — Focus
Three things I keep coming back to.
Vision–Language Models
Where grounding breaks, why VLMs hallucinate so confidently, and whether inference-time signals can catch it before a downstream system trusts it.
Post-Training
SFT, preference optimisation and RL — the part that turns a capable base model into one that behaves, refuses what it should, and stays honest about what it saw.
02 — Experience
Where the models actually shipped.
Production vision in hospitals and on vehicles, and a research stint teaching a robot to read gauges.
-
Jun 2026 — Present current
Machine Learning Engineer, Computer Vision
neo by RNT Health Insights · On-site
- Migrated GI endoscopy detection from RT-DETR to RF-DETR to get the accuracy/throughput balance that live decision support during a procedure actually demands.
- Deployed a fully on-device speech-to-report pipeline (MedASR + Qwen) on a Jetson AGX Orin — clinical documentation automated with patient data never leaving the machine.
- Engineered embedding-based frame deduplication (DINO / ResNet50 + DBSCAN) to drop redundant endoscopy frames, speeding up training and improving recall on rare anomalies.
- Optimised multimodal Vision + ASR + LLM pipelines for embedded GPUs to meet hospital latency constraints.
-
May 2025 — May 2026
Deep Learning Engineer
Flux Auto · YC S21 · On-site
- Integrated 3D LiDAR models (PointNet, PointNet++, Point Transformer) into the autonomous vehicle stack for object detection where camera-only perception failed under variable lighting.
- Designed a Progressive Morphological Filter ground-removal pipeline to separate true obstacles from road noise — preventing false positives before they reached braking decisions.
- Deployed low-light image enhancement to hold camera perception together when LiDAR data was sparse or noisy.
-
Dec 2023 — Feb 2024
Research Intern
SMSS Lab · IIT Kanpur · On-site
- Built computer vision algorithms on a robotic platform to autonomously detect and read physical gauges in hazardous field environments.
- Fine-tuned a YOLO keypoint + object detection model on a curated field dataset to localise gauge parameters in real time.
Languages
Vision & DL
3D & LiDAR
LLMs & GenAI
Edge & MLOps
03 — Selected work
Projects, and what each one was actually about.
The four builds worth showing — each links straight to the source on GitHub.
Post-training
mathguard
An RLHF-style alignment pipeline for math LLMs: harm refusal lifted from 0% to 88% with a 3.2% GSM8K alignment tax. HarmMathBench plus a DistilBERT reward model blocked all 96 jailbreak attempts against a 1.5B model.
SafetyFine-tuningEvals
Multimodal
Rasengan-XL
PEFT fine-tune of SDXL 1.0 — LoRA, QLoRA and DoRA — fitting a 2.6B-parameter model on a single 16 GB T4 with 4-bit NF4 and gradient checkpointing, so style transfer runs without enterprise GPUs.
SDXLLoRADiffusion
Systems
Socio Post Generator
PostGen: a few-shot post generator. It reads existing posts, extracts their style, tags and language, then asks an LLM to write new ones in the same voice.
Few-shotGroq APIPython
Systems
PDF2Podcast
A document-to-audio pipeline: Llama-3.2-3B rewrites a dense PDF into a two-speaker dialogue — one explains, the other asks what you'd actually ask — rendered to audio with Kokoro TTS.
Llama 3.2Kokoro TTSPython
04 — Writing
Learning in public, mostly the hard way.
23 technical writeups — things I rebuilt from scratch, papers I fought with, and the occasional note to myself. The more personal ones stay on discombobulating.
/
Nothing matches that search.
05 — Articles
Longer-form threads and write-ups.
Short essays on X, paper notes on Medium and Notion. Each one opens at the source.
X
Visual Grounding Is Front-Loaded
A causal analysis of explanations in vision–language models.
Read threadX
The Cost of Always Looking
What continuous visual attention actually costs a model.
Read threadX
The Reasoning Illusion
Why VLMs are professional liars — and why fluent explanations aren't evidence.
Read threadX
LLM Latency Profiling with vLLM
Where the milliseconds actually go when you serve a model.
Read threadMedium
The Three Pillars of Post-Training
SFT, RLHF and Constitutional AI, explained without the hand-waving.
Read on MediumNotion
Retinex-Net
Low-light enhancement, decomposed into reflectance and illumination.
Open noteNotion
Knowledge Gap Alignment
Aligning what a model knows with what it will say.
Open noteNotion
AlexNet
Krizhevsky, Sutskever & Hinton — where modern vision actually starts.
Open noteNotion
VGG-16
Simonyan & Zisserman — depth, 3×3 kernels, and a very expensive lesson.
Open note07 — Contact
Let's talk vision, post-training, or anything adjacent.
Best reached on X or GitHub. I read everything, and I answer the interesting ones.