17 end-to-end systems built and deployed across industries — each solving a real business problem.
Australian English AI Tutor
Real-time bi-directional voice AI tutor with specialized Australian accent coaching, automated pronunciation scoring (Grammar / Vocabulary / Pronunciation), and session history analytics.
▸ Eliminated dependency on human tutors · Millisecond feedback loop · Cloud-native & fully scalable
GPT-4o RealtimeSTT/TTSWebSocketsFastAPIMongoDBFlutter
AI Floor Planning Agent
Text-to-floorplan intelligent planning agent. Converts natural language room descriptions into spatially-reasoned coordinate geometry and exports production-ready AutoCAD DXF files.
▸ First draft: hours → seconds · Standardized CAD-ready output · Faster sales approval cycles
Google GeminiLangChainPydanticEzdxf
Real-Time Stock Signal Predictor
Multi-source sentiment + technical trading engine. Scrapes news & social, transcribes YouTube analysis, calculates RSI/MACD/Volume trends, and produces confidence-scored signals via custom fine-tuned LLM.
▸ Reaction latency: minutes → seconds · Emotion-free confidence-weighted signals · Multi-ticker dashboard
FinBERTVADERFaster-WhisperFastAPIPlotlyPostgreSQL
Railway OHE Measurement System
Real-time vision-based railway overhead wire monitoring. Measures stagger, height & gradient parameters using YOLOv11 + ZED stereo depth at 90 km/h with GPS tagging and tolerance alerts.
▸ 50% targeted pantograph wear reduction · No track possession required · Eliminated manual inspection
YOLOv11ZED Stereo SDKCUDAOpenCVPyQt6
AI Marketing Engine
End-to-end generative ad creative engine for e-commerce. Scrapes product pages → generates AIDA scripts → creates AI images, videos & voiceovers → delivers finished assets to AWS S3 automatically.
▸ 90% manual creative effort eliminated · Weeks → hours production · Massive A/B test scale enabled
GPT-4.1/5Fal.aiElevenLabsn8nAWS S3FastAPI
AI Fantasy Avatar Generator
Identity-preserving avatar engine built on SDXL + IP-Adapter-FaceID. Extracts facial embeddings via InsightFace, generates prompt-driven fantasy scenes while preserving face identity, with SSIM-based quality filtering.
▸ Hyper-personalized avatars · Automated QA pipeline · GPU-scalable · Near-zero manual editing
SDXLInsightFaceIP-Adapter-FaceIDFastAPIPyTorch
Automated Meeting Intelligence System
Multi-speaker transcription pipeline with noise removal, speaker diarization, timestamp alignment, and AI-generated structured executive summaries with SRT subtitle export. Integrated with Zoom.
▸ Eliminated manual note-taking · Instant structured meeting minutes · Improved decision tracking
Faster-WhisperNVIDIA NeMoDemucsGeminiPyTorch
Virtual Shoulder Measurement Tool
Camera-based shoulder measurement engine for fashion e-commerce. Uses MediaPipe Pose + Face Mesh for face-calibrated pixel-to-centimeter conversion and automatic size recommendation.
▸ Reduced return rates · Improved checkout confidence · Runs on standard consumer devices
MediaPipe PoseFace MeshOpenCVNumPy
Elevator Parts Similarity Search
Visual similarity search engine for thousands of elevator SKUs. EfficientNet-B0 extracts feature embeddings; FAISS retrieval returns ranked part matches with category prediction in seconds.
▸ Part identification: hours → seconds · Reduced downtime · Democratized expert knowledge
EfficientNet-B0FAISSPyTorchStreamlit
AI Lead Nurturing Automation
30-stage automated AI follow-up engine with reply detection, business-day-aware sequencing, CRM-based personalization, and tone progression across the full sequence — from casual to assertive to final reminder.
▸ 100% follow-up coverage · 10–15 hrs/week saved · Higher reply rates · Zero human intervention
GPT-4.1n8nGmail APIGoogle SheetsHighLevel CRM
Automated Financial Reporting System
Natural language → SQL → BigQuery execution → finance reports → QuickBooks-ready journal entries. Fully automated monthly close for multi-brand e-commerce, eliminating reconciliation errors.
▸ 95% manual workload reduction · Hours → under 1 minute · Zero reconciliation errors
GPT-4.1BigQueryFastAPIAWS S3n8n
Real-Time Talking Avatar
Lip-synced real-time avatar engine combining GPT-4o Realtime speech-to-speech with LIA diffusion-based motion synthesis and HuBERT audio feature extraction for hyper-realistic animated customer interactions.
▸ Hyper-realistic AI engagement · Fully automated animation · Real-time conversational latency
LIA DiffusionGPT-4o RealtimeHuBERTPyTorchCUDA
Civil Beam Counting Automation
PDF beam extraction and annotation engine. Uses regex + spatial proximity logic to detect beams, associate lengths, handle orientation, and produce annotated audit PDFs with a built-in verification layer.
▸ Manual hours → seconds · Eliminated counting errors · Automated audit trail
PyMuPDFPython RegexGradio
LinkedIn Research Automation Pipeline
Fully automated thought leadership pipeline: scrapes arXiv papers → parses PDFs → AI summarizes → schedules and publishes to LinkedIn. Runs 3 posts/week with zero manual drafting effort.
▸ 3 posts/week fully automated · Zero manual drafting · Consistent thought leadership positioning
Gemini 2.0 FlashFastAPIAPSchedulerLinkedIn API
AI Image Super-Resolution System
Real-ESRGAN-based 2x/4x upscaling engine for images and video, with GFPGAN face restoration, noise removal, and memory-efficient tiling — turning legacy low-res assets into 4K-ready media.
▸ 4K-ready asset generation · Seconds-level processing · Legacy media monetization
Real-ESRGANGFPGANPyTorchFFmpeg
NeuroHire — Graph-Based AI ATS
Knowledge graph-powered applicant tracking system with semantic JD matching, OCR fallback parsing, Neo4j graph modeling, and conversational database querying for natural-language candidate search.
▸ 90% screening time reduction · Semantic candidate ranking · Faster hiring cycles
Gemini 2.0 FlashNeo4jFastAPILangChain
AI Basketball Analytics System
End-to-end automated basketball video intelligence pipeline. Multi-model player/ball/net detection, persistent multi-object tracking, team classification, event recognition (shots/assists/steals/blocks), zone-aware scoring, per-player CSV stats, highlight clip generation — targeting 600k+ games/year.
▸ >85% F1 on key events (targeted) · Full analyst workflow automated · Scalable to 600k+ games/yr
DeepStreamTensorRTYOLORF-DETRPyTorchFastAPIAWS S3LangChainGemini