About
I am Mohammad Yaser Hussain, a fresh AI and Machine Learning graduate from Jayaprakash Narayan College of Engineering, affiliated to JNTU, Hyderabad.
I work across the full range of AI engineering, from systems-level thinking and performance optimization to designing intelligent pipelines and applied machine learning. I am not attached to any one approach or tool. I follow the problem and figure out what it actually needs.
Building software that works in the real world matters to me more than building software that looks good in a demo. That means caring about performance, reliability and the experience of whoever ends up using it. I am also a published researcher with peer-reviewed work in machine learning applied to real world market analysis.
Project Experience
- Windows desktop AI assistant: a local Qwen2.5-VL LLM handles text and vision reasoning, backed by a 4-block CNN (trained from scratch in PyTorch, BatchNorm + dropout) as an offline vision fallback; STT via faster-whisper runs fully local, and TTS tries Edge TTS first, falling back to Kokoro-82M and then Windows SAPI so speech still works offline
- Hybrid ML/DL intent pipeline: TF-IDF + LinearSVC (scikit-learn) handles near-instant classification and DistilBERT (ONNX Runtime) with dynamic CPU/GPU provider selection handles complex queries, both routed by a heuristic complexity score, plus a rule-based short-circuit that skips inference entirely for small talk
- Concurrent PyQt6 application decoupling DL inference in an isolated worker process, wake-word detection via openWakeWord, and OS system control with structured JSONL telemetry
- SQLite conversation history with an optional ChromaDB + sentence-transformers semantic-memory layer (feature-flagged, deterministic fallback embedder when disabled) and spaCy NLP for entity extraction, alongside real-time audio waveform and hardware telemetry panels
- Privacy-first Android app for chatting with a curated library of Qwen, Phi, SmolLM/SmolVLM, and Gemma models entirely offline, with text, vision, and audio support
- Multi-engine inference layer routing three model formats (GGUF, MediaPipe .task, and LiteRT-LM .litertlm) through dedicated llama.cpp, MediaPipe, and LiteRT-LM engines behind one chat interface
- Custom C++17/CMake JNI bridge builds llama.cpp with a Vulkan GPU backend and seven HWCAP-scored ggml CPU variants selected at runtime, gated by a device capability probe and RAM-aware load warnings that guard against freezes on low-memory hardware
- Clean Architecture with Dagger Hilt and Kotlin Coroutines; a WorkManager-backed downloader handles background model downloads and hot-swapping multi-gigabyte files while Room persists chat history without blocking the UI
- Four-stage CrewAI Flow (@start/@listen) pipelining SEC EDGAR filing retrieval, insider-transaction analysis, and week-over-week historical comparison through custom Pydantic-typed CrewAI tools, with a dedicated report-writing agent synthesizing the final output
- Dual-LLM design: a local Ollama model (Llama 3.2 by default) drafts the report, then Google Gemini 2.0 Flash independently cross-checks it against the raw transaction data and silently corrects arithmetic, attribution, and count errors before it ships
- Real-time Form 4 and 8-K ingestion via ATOM feeds with Pydantic for type-safe tool schemas and BeautifulSoup for XML parsing, plus retry-with-exponential-backoff and data-validation guardrails at every step that accumulate errors instead of aborting the run
Technical Skills
Machine Learning & Deep Learning
NLP, Speech & Vision
Generative AI & Agents
Android & Edge Deployment
Languages & Tools
GitHub Statistics
Publications
Peer-Reviewed · Journal Article · 2024
Enlightening Paths: Python's Vision into the Electric Vehicle Market
View on ResearchGate ↗Education & Certifications
Degree
Bachelor of Technology
Jayaprakash Narayan College of Engineering
Affiliated to JNTU, Hyderabad