Vitalcheffe

AEGIS: Autonomous Interceptor Engine

I spend a lot of time watching the world break on Telegram. Not for the politics, but for the raw data. After staring at geospatial anomalies on NEXUS for months, the structural flaw in modern warfare became violently obvious: the economic math of air defense is completely broken.

We are firing $4M Patriot missiles and $1.2M Aster 30s at $20,000 fiberglass drones powered by lawnmower engines. You cannot defeat a decentralized, hyper-cheap swarm with a centralized, hyper-expensive monolith. It’s economic suicide. The only way to invert that asymmetric cost curve is with a faster, smarter, and deadlier swarm. That is AEGIS.

Spectral Fusion & Byzantine Consensus

A single drone's sensor data is just noise. AEGIS fuses Sony IMX678 optical (EO), FLIR Lepton 3.5 thermal (LWIR), and Inxpect LBK-24 Doppler radar. But sensors lie, get jammed, or track flares. To solve this, I implemented distributed consensus models based on research from MIT CSAIL and Harvard.

The swarm uses Median Absolute Deviation (MAD) to instantly reject anomalous sensor data (Byzantine fault tolerance). A target only hits "CONFIRMED" status when a 2/3 network quorum mathematically agrees on its spatial centroid and thermal signature against the background sky, natively filtering out decoys.

Kinematic Prediction via UKF

Linear predictions fail completely when you're trying to intercept a maneuvering target at Mach 0.26. I implemented a 9-state Unscented Kalman Filter (UKF) relying entirely on estimation theory and sigma-point propagation developed at Imperial College London.

By propagating non-linear kinematic uncertainty, the UKF collapses the interception error margin from ±50m down to ±1.5m in under two seconds.

Full Industry Concrétisation

Writing the software is only half of the equation. I am treating this like a real defense aerospace startup. I am pushing AEGIS as close to physical reality as legally and physically possible without manufacturing a live kinetic weapon.

Beyond the Python orchestrator, I am designing the 3D CAD models of the delta-canard airframe, routing the custom PCBs for the avionics, sourcing the real-world Bill of Materials (BOM), and ensuring all structural Safety Factors (1.93) comply with MIL-STD-1522A using Toray T800H CFRP material properties. If the math doesn't clear the physical reality, the 764 N·s solid Cesaroni Pro54 booster physically cannot arm.

NEXUS: Real-Time OSINT Engine

The problem with Telegram and open-source intelligence is that following it manually is exhausting and unreliable. You're jumping between channels, cross-referencing timestamps, trying to figure out if three posts about the same explosion are three independent sources or the same rumor bouncing around. You get it wrong half the time and you don't even know when you got it wrong.

I built NEXUS to solve this. Instead of me going to the information, the information comes to me — geolocated, mathematically scored, and plotted on a 3D CesiumJS globe in real time.

Signal Aggregation

It pulls from 35+ open-source feeds simultaneously: NASA FIRMS satellite fire detections, USGS seismic sensors, ADS-B aviation transponders, AIS ship tracking, GPS jamming maps, GDELT's news archive, Wikipedia edit velocity (a surprisingly accurate early warning signal), internet shutdown detection from Cloudflare Radar, and SAR imagery from Sentinel-1.

Nothing shows up on the globe until that total crosses a strict threshold. The globe is completely silent most of the time. When something appears, it means multiple independent systems reported the same physical event within a few minutes of each other.

The Science Layer

The text analysis is grounded in published research rather than vibes. Every signal description gets scored using LDA topic modelling (Mueller & Rauh, APSR 2018) to extract the underlying topic distribution and improve semantic matching.

I also implemented the Velocity Penalty (Vosoughi et al., Science 2018) to down-weight signals that spread unusually fast relative to their content type, which is a highly reliable indicator of coordinated misinformation or CIB (Coordinated Inauthentic Behavior). Add in CUSUM sequential change-point analysis for flagging statistical volume anomalies, and NEXUS transforms raw internet noise into actionable intelligence.

Sense-Act: Sentiment Arbitrage Engine

Retail trading algorithms fail because they parse Bloomberg or Reuters headlines. By the time Reuters publishes an article about an oil refinery strike, the High-Frequency Trading (HFT) firms have already priced it in. The only alpha left for retail/independent quants is unstructured, niche domain expertise hidden in the noise.

Natural Language Processing (FinBERT)

Sense-Act is a shadow-mode sentiment arbitrage engine specifically tuned for commodities (Oil/Energy). Instead of treating all text equally, it replaces standard follower-count weighting with explicit domain expertise scoring. It uses FinBERT (Araci, 2019)—a language model pre-trained on financial communication—coupled with Siamese BERT-networks (Sentence-BERT) to perform semantic deduplication via cosine similarity.

Market Microstructure Integration

Signal generation is useless without understanding how orders hit the book. The system integrates mathematical concepts from market microstructure theory, specifically relying on Hasbrouck (1991) for measuring the information content of trades and Glosten & Milgrom (1985) for understanding bid-ask spreads in specialized markets.

To optimize the weightings of the different sentiment sources (e.g., separating an actual oil shock from macroeconomic noise a la Kilian, 2009), I implemented a Genetic Algorithm (Holland, 1992) that dynamically evolves the feed weights. Slippage and variance are calculated in real-time using Welford's online algorithm (1962), ensuring the engine never hallucinates profitability where liquidity doesn't exist.