Selected Engineering / Systems · AI · Mobile · Web
Ten production systems. One underlying philosophy: turning mathematical models, physical hardware, and human urgency into deterministic, elegant software.
" In disaster zones where cellular towers collapse and debris blankets the landscape, the first 72 hours decide survival. Phoenix X is an autonomous micro-UAV running on-device edge AI and off-grid LoRa mesh comms to locate buried victims through zero optical visibility—with zero cloud dependence.
Disaster zones share four lethal barriers: communications blackouts, zero optical visibility (dense smoke, rubble, night operations), the rapidly closing 72-hour golden window, and hazards to human rescuers from structural aftershocks.
Fuses Qualcomm Robotics RB5 compute with ThermalNet YOLOv8 multi-spectral thermal inference running at <20ms latency directly on-device. Eliminates all cloud dependencies while delivering real-time survivor heat-signature isolation.
Operates over a peer-to-peer 915 MHz LoRa RF telemetry mesh spanning 5 km line-of-sight. Relays survivor GPS telemetry, victim thermal coordinates, and UAV health packets directly to ground teams without cellular or satellite links.
Sub-2 kg AUW DGCA Micro-UAV category compliance. Custom carbon-fibre airframe with Pixhawk flight controller integration, autonomous waypoint pathing, active battery thermal throttling, and fail-safe return-to-launch (RTL) protocols.
On-device neural inference latency. 5 km peer-to-peer RF mesh range without internet, full real-time telemetry streaming to open ground station cockpit.
Real-time web cockpit deployed at phoenix-sar-uav.vercel.app rendering real-time mission telemetry, drone heading, survivor coordinates, and sensor thermal feeds.
Interactive 3D gyroscopic orientation reticle. Hover to articulate gimbal pitch & yaw kinematics; click to cycle telemetry channels.
" When an emergency strikes in an isolated area, the bottleneck is rarely a total lack of resources — it is the total lack of an intelligent coordination layer to route human help within minutes.
Critical healthcare delivery in decentralized communities suffers from severe information fragmentation. Emergency calls are often frantic, unstructured, and fail to reach nearby field responders who have the right medical supplies.
Engineered an end-to-end multi-agent dispatch network. Ingests raw audio/text distress signals, processes intent via LLM agents, tags triage severity levels, and maps tasks to verified community workers using geo-fenced proximity scoring.
Full-stack TypeScript architecture built with Next.js, reactive state pipelines, low-latency WebSocket dispatch, and Gemini API integration for adaptive multilingual emergency triage.
Prioritized zero-dependency fallback: offline SMS-burst capability when cell data is degraded, and deterministic failover routing so no distress signal remains unacknowledged.
Working full-stack deployment turning human crisis into synchronized real-world field logistics. Validated through end-to-end multi-user emergency simulations.
Proves that generative AI belongs at the translation boundary—turning chaotic, terrified human speech into strict, structured JSON payloads that existing logistics software can act upon.
" Grand Prix race strategies aren't won on fixed spreadsheets — they are won by evaluating tens of thousands of stochastic futures before the first pit window opens.
Formula 1 race engineering involves navigating extreme non-linearities: tyre compound thermal degradation, safety car probabilities, fuel-mass delta lap time penalties, and rival undercut dynamics under variable track conditions.
Engineered a vectorized Monte Carlo engine in Python leveraging NumPy arrays to execute 10,000 complete race trajectories. Incorporates non-linear wear profiles informed by FastF1 telemetry feeds.
Coupled stochastic simulation with strategy optimization algorithms to evaluate optimal pit strategy windows (lap count, compound choice, safety car response) under uncertain track conditions.
Python backend leveraging vectorized linear algebra, and an interactive React front-end delivering strategy curves and telemetry graphs.
Vectorized simulations per compute pass. Open source repository with reproducible simulation runs against historical Grand Prix telemetry datasets.
Showcases how high-performance numerical computing and clean web visual interfaces transform raw mathematical uncertainty into clear tactical decisions.
" For decades, informal auto-rickshaw fleets in India have struggled with rigged meters, currency friction, and payment disputes. AutoMeter replaces mistrust with automated verification.
Millions of urban daily commutes rely on auto-rickshaws. Analog mechanical meters are easily manipulated, lack real-time digital payment reconciliation, and force passengers and drivers into awkward disputes over change.
Engineered a dual-input tracking system combining high-frequency GPS position deltas with physical wheel rotation pulses to calculate accurate distance even through tunnels and dense urban canyons without spoofing vulnerabilities.
Upon ride termination, the system generates a dynamic on-screen Unified Payments Interface (UPI) QR code embedded with the exact calculated amount, ride transaction hash, and merchant VPA for instant zero-fee passenger checkout.
Built with Flutter and Android Native services. Designed with a sunlight-readable, high-contrast UI tailored specifically to harsh vibration, high-glare automotive environments.
Operational smart meter prototype supporting instant dynamic QR scan-and-pay, automated fare brackets, and transparent driver analytics.
Built upon my earlier award-winning Phase 1 research in transit tracking, demonstrating how software can modernize essential public mobility without imposing expensive corporate hardware.
" Most productivity tools fail because they assume unlimited human cognitive energy. Opportunity plans around your real daily energy budget and stays 100% functional without the cloud.
Cloud-dependent task managers suffer from network lag, subscription lock-in, and unrealistic planning that leads to user burnout. Opportunity was designed as an offline-first deterministic executive system.
Features an algorithm that weights task difficulty, urgency, and calendar context against a capacity-aware daily budget—preventing over-scheduling and re-prioritizing remaining tasks dynamically when schedules slip.
Crafted in pure Kotlin using Android Native SDK, Room SQLite Database for offline persistence, and WorkManager for reliable background alarms.
Strict Unidirectional Data Flow (UDF) with clean domain separation. Fully decoupled data access layers guaranteeing responsive offline UI rendering.
Zero cloud dependency. Complete local privacy, instant query responses, and resilient background task execution.
Highlights the beauty of Native Android engineering when unburdened by bloated external SDKs—delivering clean, battery-efficient tools built for human focus.
" The Chinese Remainder Theorem is usually taught through dry modular arithmetic equations. Building it as an interactive, tactile digital tumbler lock transforms an ancient theorem into intuition.
Explores the border where mathematical theorems, physical mechanical puzzles, and high-framerate front-end engineering merge. Digital recreation of a physical cardboard CRT lock mechanism.
Users rotate three modular dial rings ($m_1=3, m_2=5, m_3=7$), spin the central number wheel, and pull the sliding rod to align notched gates, release the spring latch, and swing open the safe door.
Engineered with React, TypeScript, Tailwind CSS, and Framer Motion spring physics. Accompanied by WebGL experiments in Three.js and Canvas frame-sequence animations (The Ford GT, 3D WebGL fruits, AR jewelry engines).
Pixel-perfect haptic feedback simulation, audio-visual feedback on alignment notches, and responsive touch gestures across desktop trackpads and mobile screens.
Fluid spring-physics animation running live in production. Deployed interactive web demo with open-source component architecture.
Proves that the web can feel physical, mechanical, and delightful when animations obey real inertia and physical resistance laws.
" History deserves more than a Wikipedia page. 3,100+ high-definition frames, 14 historical acts, and hardware-accelerated canvas scrollytelling — making the life of a war physician and independence commander visceral, interactive, and unforgettable.
An editorial-grade interactive digital monograph commemorating Dr. Lakshmi Sahgal (1914–2012) — physician, Commander of the Rani of Jhansi Regiment in the INA, and lifelong grassroots healer. Built to move beyond static reading with cinematic immersion driven by scroll velocity.
Hardware-accelerated HTML Canvas 2D frame sequencer scrubbing 3,100+ HD frames across 14 historical acts in real-time, synchronized to the reader's exact scroll position and velocity. Sub-pixel auto-tour mode advances at a calibrated 75px/s reading velocity.
Interactive 1940s broadsheet newspaper timeline, fullscreen archival deep-viewer for primary INA artifacts, one-click global rewatch from any act, and persistent navigation across all 14 acts. Verified against National Archives of India provenance records.
React 19, Vite 5, Tailwind CSS 3, Canvas 2D hardware acceleration — achieving 60 FPS frame scrubbing with zero dropped frames across 3,100+ assets on consumer hardware.
HD frames across 14 historical acts, deployed live in production as an archival-grade interactive documentary. Open to the public at capt-lakshmi.vercel.app.
Proves that web technology — when paired with archival rigor and hardware-accelerated rendering — can transform historical scholarship into something a reader physically inhabits rather than merely reads.
" Drop any song onto the window. Within seconds — stems separated, chords recognized by a 170-chord neural network, beats aligned to bars, and a musician-ready chord chart printed. Zero terminal commands.
Pianists, guitarists, worship musicians, and arrangers spend hours manually transcribing chord progressions from audio. Existing tools are either inaccurate, fragmented across multiple steps, or require Python expertise. Song Chord Analyzer replaces the entire MIR pipeline with a single drag-and-drop desktop app.
Employs a pretrained BTC (Bidirectional Transformer for Chord Recognition) model trained on CQT spectrograms with a 170-chord vocabulary (Major, Minor, 7th, Maj7, Min7, Dim, Aug, Sus4, Sus2, 6th etc.). Reports authentic model confidence scores and surfaces alternative candidate chords when harmonic ambiguity occurs.
Full local-first pipeline: Demucs v4 GPU stem separation (vocals/bass/drums/guitar/piano) → Bass-stem inversion detection (30–350 Hz sub-bass analysis for slash chords) → BTC chord recognition on 4 parallel harmonic channels → Beat & bar alignment → Musical section clustering. All within one VRAM-managed CUDA pass.
Engineered in collaboration with Jerin Jomy (@jerinjomy07). Jerin led the original core Windows application architecture, FastAPI orchestration, and PyTorch ML bundling (jerinjomy07/Song-Chord-Analyzer). I developed the native macOS port, Apple Silicon / MPS hardware acceleration bridges, and cross-platform packaging.
Chord vocabulary. Full local pipeline from audio file to interactive musician chord chart — with PDF/TXT/JSON export, musically-correct transposer (±11 semitones), and real-time playback with synchronized chord markers.
Demonstrates how multiple ML models (Demucs separation + BTC recognition + Librosa beat tracking + HPCP chroma templates) can be fused into a single coherent ensemble that is dramatically more accurate than any one model alone.
" Mechanical camera tracking rigs and external optical markers are cumbersome, costly, and fragile on live sets. This system is a real-time AI spatial tracking engine delivering sub-millimeter 6-DoF camera pose estimation for virtual production pipelines and live broadcast cinematography.
In virtual production stages (LED volumes and green screens), the physical studio camera must synchronize its exact 6 degrees-of-freedom (6-DoF) position and orientation with virtual cameras in Unreal Engine at 60+ FPS with near-zero latency to maintain flawless perspective parallax.
Engineered markerless optical feature tracking paired with deep neural pose estimation. Processes high-framerate optical feeds, extracts spatial keypoint geometry, and continuously computes perspective-n-point (PnP) kinematics to locate camera coordinates in 3D studio space.
Couples optical tracking vectors with high-frequency IMU telemetry via an Extended Kalman Filter (EKF). Smooths camera motion, eliminates jitter, and guarantees unbroken tracking continuity even during fast pans or momentary optical occlusions.
Outputs low-latency industry-standard FreeD and Open Sound Control (OSC) camera telemetry packets over local UDP networks directly into virtual rendering pipelines with sub-frame synchronization.
End-to-end 6-DoF tracking latency. Delivers sub-millimeter positional repeatability and sub-degree orientation precision across changing studio illumination conditions.
Replaces multi-thousand-dollar mechanical tracking hardware with an efficient computer vision algorithm running in real time—democratizing high-end virtual cinematography for modern studios.
" Classic retro gaming — Street Fighter II, Combat Cars — played against an AI opponent running real-time neural network inference at 60 FPS, directly inside classic retro game engines. Player 1 is human. Player 2 never sleeps.
A multi-platform gaming environment where a human player competes against AI neural network inference running at 60 FPS inside classic Sega Genesis emulation. The AI observes game state, predicts optimal actions, and executes them with sub-frame latency — no cloud round-trips, no API calls.
Real-time RL policy network deployed on Apple Silicon (Phase 1) with planned migration to Arduino UNO-Q edge hardware. Supports Street Fighter II (fighting) and Combat Cars (top-down racing/combat) with automatic USB/Bluetooth gamepad detection for human players.
Headless benchmark mode measures latency and throughput across 1,000+ game steps for performance regression testing. Modular core/game architecture allows new retro titles to be added as plugins without modifying the AI inference engine or emulation layer.
Python reinforcement learning environment with RetroGym/StableRetro, optimized for real-time neural inference. Firmware layer designed for Arduino UNO-Q edge deployment with constrained SRAM and flash management.
Real-time RL policy inference inside live retro emulation. Playable multi-game environment with human vs AI combat and racing, plus automated latency benchmarking harness.
Explores the edge of what constrained hardware can do when AI inference is carefully quantized, memory-managed, and co-designed with the emulation loop — treating a retro gaming platform as a serious edge AI research target.