I began with signals, control systems and embedded hardware. That origin still shows up in how I debug ML: follow the path, measure the boundary and distrust a result that has lost contact with the physical process behind it.
I’m Taha Yasin Bhatti, a machine-learning engineer and researcher based in Lahore. My early work covered computer vision, autonomous systems and edge deployment. Field operations then introduced messier material: monthly forecasts, retail geography, route planning and cloud workflows that had to run on schedule.
Care-companion robotics changed the stakes again. The language layer had to work beside vision, gait, activity and camera-based physiological signals. A weak answer in that setting affects how researchers and caregivers interpret a person’s state.
Current engineering work centres on production agents, retrieval, evaluation, privacy-aware ML and the infrastructure that keeps those systems recoverable. Public work extends the same practice into EV planning, grid integration, air quality and policy tools.
Away from a screen, I play cricket as a fast bowler. I like the uncomplicated bargain: run hard, land the ball where you meant to, and accept that the batter still gets a vote.
Memory, multimodal sensing and behavioural evaluation.
The thesis benchmarks four memory architectures: naive RAG, Mem0, A-MEM and MEM1. Evaluation uses LongMemEval and LoCoMo. The planned contribution is a reinforcement-learning controller that manages short- and long-term memory through a learned policy instead of fixed retention rules.
4 architectures · 2 benchmarks · 1 learned controller
A contamination-resistant behavioural benchmark for extended agent interaction. It measures whether memory carries a preference forward, corrects an outdated belief and changes a later decision when the history warrants it.
Behavioural tasks · Generative instances · Extended interaction
I built parts of the LLM and RAG interaction layer connecting dialogue with vision and physiological context. The wider system included ADL and gait models based on skeleton keypoints, temporal classification, camera-based rPPG estimation and detectors for agitation and fall-risk cues. The study protocol was published in JMIR Research Protocols on 4 October 2024.
2024 protocol · Multimodal sensing · Human oversight
Vision, audio and physiological signals combined for agitation and risk detection, with evaluation shaped around nursing workflows and human oversight.
3 signal families · Nursing workflow
Retrieval pipelines and graph analytics over open-source reporting for early-warning analysis related to regional stability.
RAG · Graph analytics · Open-source reporting
Tool orchestration · retrieval · evaluation · memory · structured outputs · guardrails
Forecasting · geospatial analysis · computer vision · reinforcement learning · multimodal sensing
FastAPI · PyTorch · AWS · Azure · Docker · data orchestration · edge inference