Built with purpose.
Made to be useful.

Ross Dashboard · team operations

• Separate admin and team workspaces protect the views used by administrators, managers, and recruiters.
• CV uploads help fill candidate details, including role, experience, employer, compensation, notice period, reporting line, and relocation preference.
• Client intake records company, location, contacts, vacancies, and Active or Pending status.
• Candidate flow: Pending Review → Sourced → Internal Shortlist → Client Shortlist 1 → Offer Stage. Sourcing assigns a client; manager approval or rejection requires comments, client feedback is recorded, and an offer reduces the open vacancy count.
• Signed prospects can be promoted from the pending pipeline into active client records.

Ross AI Assistant · company chat

• A floating chat tool answers questions using the company's term sheets and policies, with instructions to avoid making up information.
• The chat keeps the latest 20 messages as context. Candidate questions go to standard hiring channels; client questions about terms can be directed to the right business contacts.
• A history sidebar, typing indicator, chat bubbles, and Markdown responses support the user experience.

Address cleanup

NLTK prepares informal directions; Llama 3.3 identifies suburbs, streets, house or building names, landmarks, and direction phrases such as “behind” or “opposite.” The result is organized into a clear JSON address.

Landmark image reading

Users can add a photo of a shop, sign, or nearby landmark. The vision flow reads visible names and numbers, then checks them against business listings and OpenStreetMap points of interest.

Voice directions

Voice input supports Hindi, Marathi, and English. It removes conversational filler while preserving the local direction phrases that help locate the destination.

Live dashboard

Combines ISS position and track, solar-weather alerts, launch countdowns, local sky conditions, and learning progress. Browser location and OpenStreetMap Nominatim help resolve a user's coordinates to a city, town, or suburb.

Earth intelligence

D3.js and Canvas animate landmass data. NASA SMAP soil moisture and Sentinel-2 NDVI support regional flood-risk and crop-stress views. NASA EONET provides disaster locations, while NASA GIBS adds layers such as active fires and night lights.

Observatory and sky events

VirtualSky maps the sky using stereographic projection and local time or GPS. Astronomy calculations cover lunar phases, meteor showers, eclipses, twilight, and dark-sky conditions. Planet positions use orbital elements; conjunctions are found from angular separation.

Satellite and space weather

Satellite.js uses SGP4/SDP4 with TLE data to calculate satellite position and velocity. ISS and N2YO services provide live positions and overhead passes. NOAA data tracks X-ray flares, K-index geomagnetic activity, aurora probability, and proton-flux storm levels.

Space-weather detail

GOES-R five-minute X-ray readings group flares into A, B, C, M, and X classes. K-index data feeds geomagnetic alerts and the Zhang–Paxton auroral oval; OVATION probabilities are drawn on the globe. Proton levels at or above 10 MeV are compared with NOAA S1–S5 storm bands.

Local sensor readings

An ESP32 reads temperature and humidity from DHT11 and air-quality readings from MQ135. It sends JSON to n8n, which checks and calibrates readings before saving them in Supabase. A sky-quality score weighs cloud cover (50%), humidity (30%), and air quality (20%).

AI guide and image reading

Groq-hosted Llama 3.3 provides a conversational astronomy guide, with separate friendly and mission-control voices. Groq Whisper transcribes speech; browser speech tools can read answers aloud. Vision models analyze astronomy photos, identify objects, and return details such as distance and scale.

AI navigation and recognition

The assistant can return a hidden navigation command to open and highlight a mission. The vision flow uses Groq-hosted Llama 4 Scout, with Llama 3.2 Vision 11B or 90B alternatives, to inspect astrophotography, recognize objects such as the Orion Nebula, return distance and scale, and award 50 XP for a successful observation.

Missions, news, and discovery

Space Devs launch data uses a SpaceX API fallback. Spaceflight News includes an ISRO-priority search. Historical Apollo and Voyager records appear beside current missions; NASA APOD uses API, RSS, and page-feed fallbacks.

More discovery tools

Launch payload names such as Falcon 9 and Starlink can map to local mission art. The Discover area can also show current people in orbit and a random planet's surface gravity. APOD's fallback checks its API, RSS, then a proxy of NASA's APOD page.

Learning and progress

Cadet, Specialist, Officer, Commander, and Admiral ranks guide learning. Quizzes award 10 XP per answer; modules and image scans award 50 XP. Supabase stores profiles, missions, progress, subscriptions, and earned badges.

Learning content checks

Badges are saved to user profiles, and fuzzy matching can recognize space-station and orbital-module images. The rank path connects lessons, quizzes, and observation tasks to stored progress.

Plans and growth

The product concept includes a free Cadet plan, a Commander / Pro plan at $4.99/month with extra alerts and offline features, and a proposed classroom plan for educator access to student progress. Future ideas include telescope links and a cleaned-up space-data API. The technical plan estimates an MVP at about $75/month and describes adding Redis caching and dedicated cloud resources as use grows.

Capture and monitor

WinDivert captures incoming and outgoing packets. The monitoring layer records source and destination IP, TCP/UDP/ICMP protocol, packets per source over a sliding time window, and total packets per second.

Build a baseline

The system tracks normal packet volume over time and watches for sudden changes. Per-source packet rates help separate a busy address from unusual request bursts.

Check for anomalies

Threshold checks flag a source when its packet rate exceeds a set limit. Z-score analysis checks whether traffic differs sharply from the current baseline, helping identify distributed spikes that can evade a single fixed limit.

Mitigate and log

Suspicious packets are dropped, the source IP is added to a block list, and later requests can receive HTTP 403. The dashboard updates blocked-source counts and recent event records.

Test the flow

The development workflow moves from setup and configuration to controlled traffic testing with simulate_attack.py, then to live detection and response.

Review live activity

The Nexus Dashboard shows requests per minute in a Chart.js graph, the number of blocked targets, and an incident log with time, source address, attack label, and action status. A security feed provides additional context for operators.

Attack patterns covered

Traffic patterns include volume-based UDP and ICMP floods as well as application-level HTTP floods and Slowloris-style slow requests. Combining packet-rate limits with baseline statistics avoids depending on a fixed rate limit alone.

Designed for fast response

The report describes quick detection and low response delay compared with signature-only firewalls. Immediate filtering is intended to reduce load on the protected service while keeping ordinary traffic flowing.

Live gesture input

The React interface uses a webcam feed. It captures frames and sends them to the Flask API for processing and prediction.

Image preparation

OpenCV converts each frame from BGR to RGB, resizes it to 64×64 or 128×128, and applies Gaussian blur to reduce camera noise. Contour detection, background subtraction, and Sobel or Canny edges are optional processing steps.

Gesture classification

The report describes a scikit-learn Random Forest classifier with 100 trees. Preprocessed pixels become a one-dimensional feature vector; the predicted label is returned to the interface.

Text and speech output

Recognized gestures appear as text and can be spoken with pyttsx3 for offline use or gTTS for online speech. The audio is played through playsound.

Collect and train data

A custom OpenCV script displays a 300×300 region of interest. Keyboard input captures, converts, resizes, and saves gesture samples into labeled folders for model training.

Clear visual feedback

The dark interface uses orange accents and a hand-placement guide. During detection, hand landmarks are drawn with red points and white connecting lines, with a green bounding box, predicted label, and confidence value. Controls start detection or clear the displayed letters.