Phước Đặng — AI & Embedded Software Engineer

VNPT AI Hackathon 2025 | Leader of SAVINAI | 2nd Runner-up Prize

MeetMate SAAR - Self-Reflective Agentic RAG Engine for Meeting IntelligenceIn production rollout at LPBank

Agentic Architecture (LangGraph)
- Built a stage-aware (Pre/In/Post) LangGraph system with a single-entry StateGraph, routing by stage / sensitivity / SLA to switch realtime vs batch execution profiles
- Defined a typed MeetingState (agenda/transcript/RAG/citations/tools/metrics) to ensure cross-stage consistency, reduce re-computation, and support audit-friendly replay
RAG Systems (Enterprise, Permission-Aware)
- Implemented Self-Reflective/Corrective RAG (SAAR) with pgvector + BM25 hybrid retrieval, ACL /metadata /effective-date filtering, and controlled query rewrite + retries
- Enforced "no-source -&gt no-answer" with mandatory citations for regulated domains to reduce hallucinations and strengthen compliance
Realtime LLM Orchestration: Shipped dual-tick scheduling for in-meeting runtime: Intent/ADR (~10s) + Recap (~60s) with event-driven Q&A force-tick for instant Ask-AI responses under token/latency budgets
LLM Reliability (Evaluation & Guardrails)
- Added retrieval grading + faithfulness checks to gate outputs, trigger corrective retrieval, and fail safely when evidence is insufficient, improving answer groundedness and reducing hallucination risk
- Implemented policy-based tool gating (stage/sensitivity allow-lists + confidence gating) to prevent unsafe actions during realtime sessions
Tool-Calling & Workflow Automation
- Designed schema-first tool-calling (tasks/follow-ups/docs) with idempotency keys, human-in-the-loop approval, and auditable execution logs for enterprise integrations
Streaming Integration: Co-designed REST + WebSocket API handshake with VNPT Senior Engineers to trigger per-endpoint meeting events and stream audio -&gt ASR -&gt transcript events into the AI pipeline reliably
Production Readiness (BFSI / LPBank): Delivered a private SaaS backend deployed on VNPT Cloud for LPBank, with audit-ready logging and an extensible stack for PostgreSQL OLTP + vector store + RAG hub + multi-LLM providers

Link github: ...
Link certification: ...
Link Product MVP: ...

View case study

SCENT — System for Customer EXperience, iNventory & Threats

End-to-end AIoT system spanning firmware, gateway, software, hardware, and AI platforms.

Thread/BLE/UART/I2C/MQTT dataflows on EFR32 + Raspberry Pi.
Flask APIs + PostgreSQL + Redis with analytics-ready pipelines.
1st runner up (SILABS IoT Challenge).

Designed & delivered a resilient end-to-end AIoT system spanning firmware, gateway, software, hardware, and AI platforms.

System design & integration: Architected a 24/7 AIoT system with robust dataflows over Thread, BLE, UART, I²C, MQTT.
MCU firmware (EFR32):
- Developed HX711 load-cell driver & IR sensor interrupts.
- MicriumOS tasks:
  - Glass-break detection (I2S mic, 200 ms loop).
  - Temp/humidity sensing (SI7021, every 5s).
  - Continuous OpenThread networking.
- Enabled Thread-to-gateway messaging with failover.
Edge gateway (Raspberry Pi 4):
- Configured as OTBR (MG21 RCP + Spinel).
- Python scripts for Thread parsing, UART, I²C LCD1602.
- MQTT gateway + PostgreSQL schema & ETL scripts.
- Local DB for scent notes & shelf ops with sync/offline mode.
Software & data platform:
- Full-stack Flask app (APIs, templates, admin).
- Smart Screen UI/UX (HTML, CSS, JS).
- Backend with Redis + PostgreSQL; Python workers for sync.
- Collected & processed 5k+ interactions for analytics & retraining.
Hardware engineering:
- Integrated sensors (load cell, cam, mic) with MCUs & Pi.
- PCB soldering, wiring validation, shelf-mounting design.
AI:
- Python pipelines for dataset generation & labeling.
- NLP pipeline (BERT + spaCy NER) for keyword extraction.

Awards: 1st runner up (SILABS IoT Challenge) — Granted Internship Certification Github Link: https:// ...

View case study

AIMING - AIoT Infravision for Agricultural Quality

Edge AIoT system on Intel hardware ensuring produce quality via dual AI modules.

• Hardware: Intel® NUC (CPU+GPU), 720p camera.
• Software: OpenVINO™, Intel® Tiber™ Cloud, Edge Software Hub.
• AI Modules:
– Vision AI: fruit type, count, ripeness, external defects.
– NIR AI: °Brix, moisture, internal bruises, pest/disease detection.
• Evaluation: Real-time grading, defect detection, SDG-aligned impact.

Architected and delivered an edge AIoT grading system on an Intel® industrial PC (CPU/GPU), integrating NIR spectral sensing + metadata pipeline for on-site agricultural quality assessment.
Built an end-to-end ML pipeline in Python: feature engineering from 6-channel NIR spectra (610-860nm) + fruit-type one-hot + ripeness (10-dim input), scaling (StandardScaler), train/validation split, and evaluation with regression + classification metrics.
Multi-task deep learning model (TensorFlow/Keras): Multi-branch 1D CNN residual blocks (kernel 2 & 5) + positional embedding + 2-layer Multi-Head Self-Attention + attention pooling + shared MLP with multi-head outputs:
- Regression: °Brix & Moisture
- Classification: Grade (A/B/C), Defect (Y/N), Fungus (Y/N)
Training included Huber loss (regression) + focal losses (classification), loss weighting, and EarlyStopping for stability.
Productionized inference for Intel hardware: exported Keras to ONNX (tf2onnx) and converted ONNX/OpenVINO to OpenVINO IR (.xml/.bin), achieving up to ~3x faster inference on Intel® devices.
Built a real-time monitoring dashboard (frontend + backend) with MQTT + Redis to stream and visualize grading outputs, enabling live operational monitoring and faster on-site decision-making.

View case study