LAYER        STACK                                          STATE
  backend    FastAPI · Node.js · async SQLAlchemy           ● active
  frontend   React · Vite · Tailwind · PWA                  ● active
  data       PostgreSQL · PostGIS · Supabase · MongoDB      ● active
  security   Ed25519 · X25519 · ChaCha20-Poly1305 · JWT     ● active
  infra      Docker · Vercel · Render · GitHub Actions      ● active
  orchestr.  Kubernetes                                     ○ loading…

_{The security row isn't decoration — Pukar signs 15 immutable fields per SOS packet and seals the payload so relays forward what they cannot read.}

Hypothesis — a market is not a set of independent tickers; it's a network, and a shock propagates along its edges.

A system exploring how a sudden movement in one stock can influence related stocks — turning the market into a network of cascading events instead of isolated price series.

Python Machine Learning Data Analysis Time Series

STATUS: 🟢 in progress

Hypothesis — disaster response fails at the seam between prediction and delivery. Build across the seam.

DETECT ──▶ UNDERSTAND ──▶ PREDICT ──▶ ASSESS RISK ──▶ WARN ──▶ DELIVER ──▶ SURVIVE NET FAILURE

PyTorch FastAPI Kotlin/Android Groq

🔗 RealArnav007/CHAKRAVYOOH

Hypothesis — during an incident, almost every alert is a symptom. Find the one that isn't.

Alert correlation & deduplication engine — Team ZenVerse @ Synergy 2026, HPE Problem Statement #10. It ingests the raw stream and in real time correlates temporally + semantically related alerts, ranks the likely root cause with a confidence score from topology/timing/severity, suppresses derivative noise, and summarizes each incident into a one-paragraph brief with a recommended first action.

READOUT · benchmark aiops-scn1 (labeled ground truth)
  Root-cause Hit@1  ████████████████████░   92.3%
  Root-cause Hit@3  █████████████████████   100%
  Cluster purity    █████████████████████   100%

_{Reproducible via backend/eval/harness.py — numbers, not vibes.}

FastAPI Embeddings React LLM

🔗 ZenVerse-synergy-2026

Hypothesis — most Indian emergency patients never touch an ambulance. Design for the family already in the car.

A PWA + feature-phone SMS service for the self-transporting emergency family. One POST /emergency fans out into two simultaneous actions:

1. Hospital — rank by department + proximity, send one-tap confirmation links; the first hospital to tap Accept takes the patient. A human confirming a bed, never a stale number.
2. Blood — match compatible nearby replacement donors and route them to the nearest licensed blood bank.

Three-layer backend with InMemoryStore / SupabaseStore behind one get_store(): the demo runs with zero external services, production is a true drop-in. 31 tests passing, CI on every PR, API_CONTRACT.md as the single source of truth for both sides.

FastAPI Supabase/PostGIS React 19 PWA

🔗 SarmaHighOnCode/GoldenHour · Bharat Academix CodeQuest 2026

Hypothesis — the network is the first casualty. Assume it's gone and design the protocol anyway.

Phone = communicate. Backend = understand. Web = act.

graph LR
    V[Victim device<br/>signs + seals] -.->|BLE / Wi-Fi Aware| R[Relay nodes<br/>forward blindly]
    R -.-> G[Gateway device<br/>has internet]
    G -->|POST /sos/ingest| API[FastAPI ingestion]
    API --> C[Crypto verifier<br/>Ed25519 + X25519]
    C --> ML[AI triage<br/>Groq Llama 3]
    ML --> Z[Geo-spatial zone engine]
    Z --> D[Commander dashboard<br/>realtime WebSocket]

• Zero-trust by construction — 15 immutable fields serialized to canonical bytes and Ed25519-signed; any relay tampering returns 401 BAD_SIGNATURE. Payloads are X25519-sealed (ChaCha20-Poly1305), so relays carry what they cannot read.
• Correlation engine — Haversine clustering folds reports within 500 m / 2 h into one incident; zones escalate NORMAL → EMERGING → HIGH → CRITICAL → EXTREME.
• Replay-proof — ±5 min clock-drift window + DB-backed unique msg_id.
• Tested — 18-suite pytest integration layer, containerized for Render.

FastAPI Ed25519/X25519 PostgreSQL Next.js

_{Private repository — walkthrough available on request.}

Hypothesis — an unexplained fraud score is an unusable fraud score.

Fuses a Random Forest classifier with NetworkX graph analytics to surface money-laundering rings — then explains why each transaction was flagged, so an analyst can act on it instead of trusting a black box.

FastAPI scikit-learn NetworkX React

🔗 Money-Trail-Engine

CatBoost/LightGBM ensemble predicting per-match goal-scoring probability, validated out-of-fold (AP 0.45) with SHAP for feature attribution. Built for the IEEE CS MUJ datathon and shipped as a Streamlit app.

CatBoost LightGBM SHAP Streamlit

🔗 offside-football-prediction

Finding — theory and a real motor disagree, and the motor is right.

Line-following robots, sensor arrays, motor control, PID tuning, ESP32 systems and autonomous behaviour — including a from-scratch PID line-follower on ESP32 + TB6612FNG with a 16-array IR sensor bar and a custom 3D-printed chassis. Control loop written and tuned by hand.

Now extending into ROS and autonomous systems during the VimaanX residency.

ESP32 C/C++ Embedded Systems ROS

Hazard analytics, driving rules for 120+ countries, India/USA fine calculators and emergency SOS — all working with no connection. Built for the National Road Safety Hackathon @ IIT Madras.

FastAPI React TypeScript PWA

🔗 roadguard-ai