AI / ML engineer with experience across NLP, computer vision, signal processing and time-series sensor data. Production ML at Sahha (behavioural intelligence from wearables), NLP pipelines at Hypadrive, and a personal portfolio spanning RF / SDR, MediaPipe-based vision, custom-trained ONNX demographic models, edge LLM deployment (30B MoE on iGPU), AI medical triage, LLM integration and adversarial-content detection.
Technical Skills
Languages
Python (primary) · JavaScript / TypeScript · SQL · Kotlin / Java (Android) · Familiarity with C / C++ and Go
MediaPipe-based gesture / hand / face tracking pipelines · Servo control (FEETECH ST-3215-C047) on InMoov i2 robotic hand · SO-ARM101 robot arm with MuJoCo simulation + imitation learning · 3D printing for drones, cinewhoops, payload grabbers, perching legs
Real-time computer-vision pipeline for smart-glasses use cases: MediaPipe Face Mesh (468 landmarks), ArcFace recognition, multi-face tracking at 30fps with a persistent identity library — all in-browser. Ships custom-trained 'Folkus' demographic models (2.3 MB INT8 ONNX, running via ONNX Runtime Web) that beat commercial APIs: gender 92.1%, age 5.37y MAE, ethnicity 72.2% across 7 groups. Camera glasses stream to a host computer that recognises faces and feeds AI context back via phone.
Multi-layer software-defined-radio intelligence stack: captures RF across UHF/VHF/HF/marine/aviation bands, demodulates in real time (WFM/NFM/AM/SSB), transcribes voice traffic with faster-whisper on AMD ROCm, extracts entities (callsigns, locations) and archives everything into a searchable SQLite/FTS5 database. The DSP core was extracted into a standalone library (rfai-dsp) with anti-alias decimation, phase-continuous VFO mixing and an abstract SDR backend.
30B-param MoE LLM running 100% on an AMD iGPU — 38 tok/s, zero cloud
Deployed Z.ai's GLM-4.7-Flash (30B-A3B Mixture-of-Experts, Q4_K_M, ~19 GB) entirely on an AMD Radeon 8060S iGPU via ROCm 7.2. The default 198K-context model OOM'd (KV cache alone needed ~53 GiB vs 32 GiB VRAM); solved by engineering a custom 32K-context variant that brings total memory to ~28 GiB for 100% GPU execution with zero CPU offload. Integrated as the offline fallback for local coding agents.
Heart rate, HRV and SpO₂ from a webcam — 100% in-browser rPPG
Browser-based remote photoplethysmography pipeline extracting heart rate, HRV, SpO₂ and stress indicators from regular webcam video. No wearable required — pure signal processing on a noisy real-world input.
Stack: JavaScript · rPPG · Signal processing · Browser ML
Full-stack intelligence pipeline that continuously ingests public message boards, runs NLP for topic clustering, sentiment, authenticity and signal detection, flags narrative spikes via z-score detection, and auto-generates balanced, sourced journalistic articles with SEO/OG metadata. Live dashboard with rolling 24h/7d/30d windows; 5,400+ story artifacts and 14 auto-published articles to date.
AI medical triage system using conversational LLMs to capture patient symptoms and route to appropriate clinical pathways. Built around the same behavioural-data thinking as the Sahha role.
