Mahin Nandipa mahin-aeroai

About Me

class MahinNandipa:
    role      = "AI/ML Engineer × Aerospace Systems"
    flagship  = "ORBITIQ-X — Aerospace Intelligence Platform"
    what      = """
        Vertically integrated reasoning stack for Earth orbit.
        Couples orbital mechanics · knowledge graphs ·
        agentic AI · digital twins.
        SGP4 -> Neo4j KG -> BGE-M3 RAG -> LangGraph 7-agent.
        50K-object screening · Foster Pc · re-entry monitor.
    """
    education = ["B.Tech Aerospace Eng · VIT Bhopal · CGPA 8.23",
                 "PG Cert AI/ML · IIIT Hyderabad",
                 "TalentSprint × Accenture"]
    location  = "Hyderabad, India 🇮🇳"
    past      = ["Acoustic anti-drone detection hardware",
                 "@ Dautya Aerospace (11 months)"]
    seeking   = ["GenAI / MLOps roles",
                 "Aerospace AI research",
                 "Defense-adjacent autonomy"]
    focus     = "Orbital mechanics + deep learning + agentic AI"
    measured  = ["+34pp CodeBLEU (38% -> 72%) · LoRA r=16 · FAISS",
                 "F1=0.984 LSTM autoencoder on telemetry",
                 "50K-object conjunction screener · Foster Pc",
                 "3-layer grounding guard (NLI + numeric + flags)"]

🌌 ORBITIQ-X — Aerospace Intelligence Platform

A vertically integrated AI reasoning stack for Earth orbit. Not a dashboard, not a chatbot — orbital mechanics, a knowledge graph, multi-agent retrieval and a live digital twin wired into one coherent system.

What ORBITIQ-X is:

🧠 Purpose-built for space operations — the reasoning layer, ontology and safety gates are designed around orbital mechanics rather than retrofitted onto a general-purpose assistant
🛰️ Digital Twin of Earth Orbit — 50K+ tracked objects, SGP4/SDP4 propagation, live TLE feeds (2h refresh), conjunction risk layer, orbital density heatmap
🔗 Aerospace Knowledge Graph — Neo4j, 13 node types (Satellite · Orbit · Mission · LaunchVehicle · Conjunction · Debris · SpaceWeatherEvent · ResearchPaper + 5 more), 40+ relationship types
📚 Physics-grounded RAG — BGE-M3 1024-dim dense + SPLADE sparse hybrid search over a NASA/ESA/ISRO corpus, HyDE query expansion, 3-layer grounding guard (NLI entailment + numeric verification + uncertainty surfacing)
🤖 7-Agent LangGraph System — Orbital Dynamics · Conjunction Analysis · Space Debris · Mission Planning · Space Weather · Aerospace Research · Satellite Intelligence — running in parallel over a shared typed state behind a safety gate
🛡️ Safety-first design — maneuver gate (ΔV > 10 m/s requires human sign-off), Pc ≥ 1e-3 external alert, re-entry escalation, space weather gate (Kp > 5)

Architecture at a glance

Layer	Technology
Orbital engine	python-sgp4 · Poliastro · SciPy DOP853
Knowledge graph	Neo4j 5.x · GDS · 13-node ontology
RAG system	BGE-M3 · Qdrant · BGE-Reranker-v2
Agents	LangGraph · Claude Sonnet · 7 agents
Safety	Custom SafetyGate · 5 control layers
Frontend	Next.js 14 · Cesium/Resium · FastAPI
Storage	PostgreSQL · Redis · InfluxDB · MinIO
DevOps	Docker Compose · Prometheus · Grafana

Roadmap targets (design goals — not yet measured)

Metric	Target
Tracked objects	50K → 100K
KG nodes	10M+
RAG faithfulness	≥ 0.90
Screening cycle	15min → real-time
Agent response	< 3s (GPU)

python-sgp4 Neo4j Qdrant BGE-M3 LangGraph Claude FastAPI Next.js Cesium

🔬 ORBITIQ-X Technical Depth — click to expand

Orbital Engine — SGP4/SDP4 propagator with J2 correction, WGS-84 coordinate transforms (TEME ↔ ECI ↔ ECEF ↔ geodetic ↔ LVLH), orbit classifier (VLEO/LEO/SSO/MEO/GEO/HEO/GTO), two-stage conjunction screener (voxel-hash spatial prefilter → Foster 2001 Pc), Clohessy-Wiltshire relative motion with two-impulse rendezvous, re-entry monitor (King-Hele drag + NRLMSISE-00 density lookup), TLE parser with checksum and field validation.

Aerospace RAG — 7 source types (PDF/HTML/LaTeX/DOCX): NASA mission reports, ESA documents, ISRO technical reports, research papers, debris studies, orbital mechanics textbooks, operator manuals. Aerospace-aware chunker (equations and tables kept as standalone chunks, sentence-boundary protection for "Eq. (3.14)", "Fig. 4"). BGE-M3 1024-dim dense + SPLADE sparse named vectors in Qdrant. HyDE hypothetical-document expansion. Grounding controls: DeBERTa NLI entailment, ±10% numeric verification, [UNCERTAIN] marker surfacing. IEEE inline citations with full bibliography.

LangGraph Agents — OrbitalState TypedDict with annotated list reducers for parallel-safe writes. Supervisor: Claude-based 10-category intent classifier → deterministic ROUTING_TABLE. 28 tools across 7 agents, all async and timeout-bounded. Safety gate: 5 layers (maneuver · re-entry escalation · collision escalation · space weather · data quality). SSE streaming for live agent activity.

Knowledge Graph — All constraint and index DDL idempotent (IF NOT EXISTS). Three index strategies: range (altitude, Pc, dates), fulltext (names, abstracts), and vector (1024-dim cosine HNSW, matching BGE-M3 output, for GraphRAG ANN). GraphRAG: hybrid Cypher + vector retrieval merged into a single context block. 14 example aerospace-intelligence Cypher queries including GDS PageRank and community detection.

🛰️ ORBITIQ Sentinel V3 — Mission Ops Center

Autonomous mission operations center — LSTM autoencoder anomaly detection on spacecraft telemetry, live orbital map, SHAP root-cause analysis, ranked maintenance recommendations.

Capabilities:

🔴 Live orbital map — Earth with LEO/MEO/GEO shells, active satellites, debris, threat objects
🧠 LSTM autoencoder — 18 telemetry parameters, reconstruction-error anomaly scoring
🛸 Digital twin — spacecraft schematic with per-subsystem health and RUL
⚠️ Conjunction risk — CDM cards with TCA, miss distance, Pc, maneuver recommendations
🔮 Predictive maintenance — component failure probability and days-to-failure
🧬 Root-cause SHAP — causal chain from solar storm → battery degradation
🤖 Ranked recommendations — action plans with confidence and estimated risk reduction

Model performance

Metric	Score
F1	0.984
Precision	1.000
Recall	0.968
PR-AUC	1.000

_{Evaluation setup: 18 telemetry parameters · held-out test split}

PyTorch LSTM SHAP SGP4 Chart.js

🌌 SSA Trilogy — Space Situational Awareness Platforms

Three interconnected browser-native SSA platforms. All open-source, all live.

🔵 ORBITIQ Explorer v2

15 modules · SGP4 · AI Copilot

ISS/Starlink/NavIC · Conjunction screening · NOAA space weather · NASA NeoWS · 3D globe · Pass predictor (Hyderabad) · Claude AI Copilot

🟣 ORBITIQ NEXUS

Apex SSA · Heatmap ring · Knowledge Graph

9 modules · Electric indigo design · Live orbital density heatmap · Cloudflare Worker API proxy · Knowledge Graph canvas

🟡 ASTRA WATCH

Ops monitoring · ISS live · Dashboards

8 modules · Electric amber on ops navy · Live ISS position · Space weather · Debris density dashboards

🏗️ All Projects

Full project index — 13 projects

Aerospace & SSA

Project	Description	Live
🌌 ORBITIQ-X	Aerospace intelligence platform · 7-layer stack · SGP4 orbital engine · Neo4j KG (13 node types) · BGE-M3 RAG (Qdrant hybrid) · LangGraph 7-agent system · Safety gate · Digital Twin of Earth Orbit · 50K object capacity	—
🛰️ ORBITIQ Sentinel V3	Mission ops center · LSTM autoencoder F1=0.984 · Orbital map · Digital twin · Conjunction risk · Predictive maintenance · Root-cause SHAP
🚀 Rocket GNC — Trajectory Correction	Planar 3-DOF flight simulator · RK4 integration · PID attitude controller · Bayesian-optimized gains · ~52% RMS attitude-error reduction · Brahmastra capstone (19-page report)	—
🌌 ORBITAL NEXUS SSA	Interactive SSA infographic (companion to ORBITIQ NEXUS) · 10 sections · Full ops stack from sensors → SGP4 → Alfano/Foster Pc → maneuver decisions
🌐 Space Data Intelligence Hub	24-module educational SPA · Two-phase architecture · Inline SVGs · Orbital mechanics through mission design
🛡️ Anti-Drone PINN	Physics-informed neural network · 2D wave equation · Fourier features · FFT/STFT · LSTM anomaly detection · Streamlit dashboard

AI / ML

Project	Description	Live	Repo
🧠 CodeGen RAG	+34pp CodeBLEU (38% → 72%) · LoRA r=16 on codegen-350M-multi · FAISS retrieval · IIIT-H PGCP capstone (Group 39)	—
🤖 BI Copilot	NL → SQL → chart → insight · Schema-aware RAG over table/column metadata · FastAPI + Streamlit	—	Private

Data Systems & Tooling

Project	Description	Live	Repo
🏢 MMDI ONE	AI-native enterprise platform · Next.js 16 · React 19 · TypeScript · Tailwind v4 · Supabase · 42-component design system with tokenised theming · 31 workspace modules · AI Copilot with 16 grounded tools · bin-packing signage cost estimator · 333-document archive on Cloudflare R2
🗄️ AeroIntel DB	603-platform India aerospace intelligence database · Wikipedia REST API · offline AI fallback · single-file HTML
📚 GDPPE	Global Database of Production Printing Equipment · provenance-first schema — every spec traces to a source fetched in-session · 9 categories, 44 machines, 149 sources, 33 logged conflicts
🏭 AeroLens	India aerospace ERP · 87 companies · .NET 8 API (60+ endpoints) · Bootstrap 5 · Oracle DDL
🧾 SignERP v2	Single-file browser ERP · soft signage & LED fabrication · costing, quoting, production management