Alt+P Sys_Properties.cpl │ Alt+S Core_Engines.exe │ Alt+W WinAmp_Audio.exe │ Alt+D Device_Manager.msc │ Alt+M Sys_Monitor.exe │ Alt+C Outlook_Express.exe │ Win+E Portfolio.url │ Ctrl+L LinkedIn.url
I build distributed systems and applied deep learning models: high-throughput extraction clusters, computer vision phoneme classifiers, and webhook servers that handle strict sub-200ms acknowledgement windows.
C:\PORAS> type status.sys
+ OPERATOR : Poras Nagar
+ ROLE : Full-Stack AI Engineer & Distributed Systems Developer
+ DEGREE : B.Tech CSE (Hons. AI & ML) — Amity University [2021-2025]
+ LOCATION : Noida, India (UTC+5:30) • Open to Global AI & Backend Roles
+ CORE FOCUS : Computer Vision (Lip-Reading), Distributed Queues & Low-Latency APIs
+ AVAILABILITY: Available for Hire / Full-Time Engineering TeamsC:\PORAS> tree /f
C:\PORAS
├── APPS
│ ├── LUMAVOICE.EXE [Hindi Lip-Reading Computer Vision Model]
│ ├── TRIAGE_GATEWAY.SYS [Hospital WhatsApp Meta Cloud API Webhook]
│ ├── SCRAPER_BROKER.BIN [Distributed 10,480 pg/min Extraction Queue]
│ └── HERMES_AGENT.EXE [Multi-Agent Autonomous CI Document Pipeline]
├── DRIVERS
│ ├── PYTORCH.SYS [TorchVision / CUDA / Spatio-Temporal CNN-LSTM]
│ ├── OPENCV.DLL [Facial Landmark / Mouth ROI Extraction]
│ ├── FASTAPI.EXE [Async Python Microservices & REST Endpoints]
│ ├── REDIS_CACHE.SYS [In-Memory Fingerprinting & Deduplication]
│ └── RABBITMQ.BIN [AMQP 0-9-1 Topic Exchange Queue Broker]
└── STATUS.BAT
Operator: Poras Nagar | Status: ONLINE | Ready for Interviews
Four production systems & deep learning architectures I'd be happy to be interviewed about line by line.
A spatio-temporal deep learning model that decodes spoken Hindi directly from silent video streams by analyzing lip motion dynamics and mouth ROI frames.
|
Patients book slots, submit intake and receive pathology reports without leaving WhatsApp — no app install, which is the whole point in a tier-2 hospital's catchment area.
|
The ingestion layer everything else sits on. A distributed extraction cluster that pulls high-frequency datasets, filings and registrar data, then normalises it into one schema before anything downstream sees it.
|
Agentic tooling deployed across developer repos: it drafts documents, generates decks and spreadsheets, and takes the repetitive half of engineering paperwork off the team's plate.
|
Prior Engineering Track Record: Architected high-throughput OTC pricing engine & B-tree query caching at UnlistedStox (sub-15ms reads).
Interactive retro audio player with live spectrum equalizer and channel telemetry.
📁 C:\PORAS\RESEARCH\LUMAVOICE.DOC — Hindi Lip-Reading Model Architecture
- Input Stream: Raw 1080p 30 FPS silent video stream capturing oral cavity motion.
- Facial ROI Cropping: Dlib 68 facial landmark detector isolates coordinates [48:68] (lip contour), normalized to 112×112 greyscale tensor.
- Front-End Feature Extraction: 3D Spatio-Temporal Convolution (Conv3D, kernel 3×5×5) captures localized micro-movements across temporal windows.
- Sequence Modeling: 2-Layer Bidirectional LSTM (512 hidden units each) resolves temporal dependencies across phonetic Hindi transitions.
- Output Decoding: CTC (Connectionist Temporal Classification) beam search decoder maps phoneme sequences to Hindi Devanagari script representation.
Repository: github.com/porasnagar/LumaVoice
📁 C:\PORAS\SYSTEMS\TRIAGE_DISPATCH.SYS — Hospital WhatsApp Cloud API Architecture
- Verification Gate: Express.js middleware computes HMAC-SHA256 signature against Meta App Secret and acknowledges incoming POST with
200 OKwithin 142ms.- Asynchronous Task Queue: Incoming patient payloads immediately pushed to Redis buffer, freeing the HTTP event loop.
- State Machine: Multi-turn triage questionnaire state tracked in MongoDB session store with 15-minute sliding TTL.
- Report Pipeline: Automated pathology PDF download links generated with 256-bit encrypted pre-signed expiring tokens.
📁 C:\PORAS\SYSTEMS\SCRAPER_CLUSTER.BIN — Distributed High-Throughput Scraper Cluster
- Broker: RabbitMQ AMQP 0-9-1 topic exchange with separate priority queues for filings, order depth, and registrar tables.
- Deduplication: Redis SHA-256 content hashes with sliding 48-hour expiration to eliminate redundant extraction workloads.
- Dead-Letter Replay: Failed parses automatically routed to DLX with exponential backoff and jittered retry intervals.
- Target Schema: Strictly validated Pydantic models at the queue boundary before storage in PostgreSQL.
📁 C:\PORAS\SYSTEMS\HERMES_AGENT.EXE — Multi-Agent Autonomous CI Pipeline
- Engine: Autonomous task agents running in GitHub Actions triggered by cron schedules or repository events.
- Artifact Generation: Deterministic rendering pipeline producing formatted DOCX, PPTX presentations, and XLSX financial models.
- Version Control: Generated output automatically validated, tagged, and committed back to dedicated artifact branches.
Windows 98 Device Manager / CRT Process Monitor showing verified technology drivers.
Expand Complete Driver Registry
| AI & ML Models |
|
| Backend & Queues |
|
| Databases |
|
| Infrastructure |
|
Live telemetry stream refreshed automatically via GitHub Actions CI pipeline.
Disk Defragmenter / Contribution Calendar (3D)
Snake Eats the Cluster Sectors
Dual Retro OS Architecture (Microsoft Windows 98 Light Mode & 90s Phosphor CRT Dark Mode) • Hand-crafted SVG Design System • Auto-updated via GitHub Actions CI.