Harsh Pandhe is a robotics and systems engineer based in Navi Mumbai, India. He builds autonomous navigation stacks (ROS 2, PX4, LiDAR SLAM), multi-agent reinforcement learning for robot swarms, quantitative trading research, and geospatial infrastructure software. He is a Software Development Intern at Doppelmayr, a Finalist and Automation & Navigation Lead at the ISRO Robotics Challenge (IRoC-U 2026), the Smart India Hackathon 2025 National Winner, and an SSRN-published quantitative finance researcher.
Teensy firmware, a Rust order-matching engine, Sentinel-2 pipelines, ROS 2 swarms, Solidity contracts. 140+ repositories in, the range is less scattered than it looks — I keep landing in the same place: systems that touch the physical world, where being wrong is expensive and nobody grades you on the demo.
name: Harsh Pandhe
pronouns: he/him
role: Former Software Development Intern, Doppelmayr (Apr – Oct 2026)
focus: Autonomous navigation · Multi-agent robotics · Quantitative research · Geospatial infrastructure
location: Navi Mumbai, Maharashtra, India
education: B.Tech Computer Science, Sinhgad Institute of Technology (SIT) Lonavala, 2024–2027
awards:
- Smart India Hackathon 2025 — National Winner (Team Algo Sapiens, 10,000+ teams)
- ISRO Robotics Challenge IRoC-U 2026 — Finalist, Automation & Navigation Lead (Team ID 10144)
- RoboRashtra 2026 — 1st Runner-Up, Chakravyuh maze & line solving
- Singapore–India Hackathon 2026 — selected for the Grand Finale (SMU, 15–17 Nov 2026)
publications:
- SSRN (2026) — "When Gradient Boosting Does Not Beat Momentum" (quantitative finance)
- SSRN (2026) — "The Cost of Sector Constraints in Cross-Sectional Momentum"
- SSRN (2026) — "Hysteresis Rebalancing in Cross-Sectional Momentum: Reducing Boundary Churn and Transaction-Cost Drag in Indian Equities"
- SSRN (2026) — "MARS: A Safety-Filtered Multi-Robot Area-Coverage Testbed with Classical, Random-Walk and MAPPO Controllers"
- SSRN (2026) — "A Failed Diagnosis of Entropy Collapse in Small-Budget MAPPO for Multi-Robot Area Coverage"
- SSRN (2026) — "Diagnosing Low-Displacement Failure Mode in MAPPO for Multi-Robot Area Coverage"
- Zenodo (2026) — MARS multi-robot area-coverage testbed, DOI 10.5281/zenodo.22758285
- IRJMETS (2024) — "Collaborative Real-Time Code" (distributed systems)
leadership: Web, Cybersecurity & Blockchain Head, IIC-SIT · Lead Organizer, Project Morpheus 2026
core_stack: [ROS 2, PX4, Gazebo, Python, PyTorch, Rust, TypeScript, Next.js, PostgreSQL, Solidity, Docker]
open_source: 101 public repositories · open PRs to validator.js and Textualize/rich
contact: harshpandhehome@gmail.com · https://harshpandhe.com · https://github.com/harsh-pandhe
open_to: Research collaborations · Open-source work · Robotics & quant rolesNavigate: Currently · Research · Robotics · Quant · Geospatial · Full-stack & AI · How I work · Experience · Stack · All projects · Contact
- 🎆 Taramandal is pushing down the V1 single-drone path. ~20 PRs merged with the contributor team in the window: a single-drone operating mode with telemetry-freshness checks and staged hover progression, a fail-closed export gate that blocks any show whose validation is unverified or failed, the 120 m altitude ceiling enforced at every backend gate, exports re-certified after centimetre rounding (so what's certified is what flies), intermediate-morph separation checks with altitude layering, and every status endpoint and audit entry labelled with its simulation backend and fidelity tier. Studio gained a safe SVG-to-formation importer, split-formation and depth-copy primitives, and typed refusals instead of silent failures. Sixteen new issues filed against my own stack — including the 500-drone validation latency — so the gaps are on the record.
- 💼 Doppelmayr — DoppelDash moved from feature work to launch readiness: phased staging verification with evidence ledgers, a read-only production review and cutover plan, a DB-outage fail-closed wrapper and lease-based cron claims, SSRF policy on user-configured mail hosts, purpose-bound private documents, a Next 16 / React 19 upgrade, and ~40 routes hardened to return 400 instead of 500 on bad input — web CI now at 1,071 passing tests, alongside the Flutter Android client heading to the Play Store. Still inside confidential projects valued in the millions, against strict international engineering standards.
- 🤖 MARS v1.0.0 stands as a research artifact — companion write-ups rebuilt on archived data, including a failure-to-reproduce report backed by a controlled A/B test; archived on Zenodo (DOI 10.5281/zenodo.22758286).
- 📊 AlgoTrade — the F&O research track stands on date-effective instrument, cost and margin models (30–45 invariant tests each); the first candidate, OPT-FLOW-001, came out of its falsification battery INCONCLUSIVE → HOLD and stays there until the evidence says otherwise. The real-money live pilot keeps running beside the paper book.
- 🛰️ IRoC-U 2026 & a new CubeSat track — the IRoC sim stack now rebuilds on a fresh machine from a single script, after the arena closed at 30/30 autonomy runs. Separately, a 1U CubeSat engineering effort kicked off with a first sprint split across systems, flight software, electronics, mechanical and ground segment.
↻ Updated 2026-10-29
A Significance-Tested, Walk-Forward Negative Result on Cross-Sectional Stock Selection in the Nifty 500 · SSRN, 2026 · independent researcher
An XGBoost cross-sectional ranker is tested head-to-head against a plain 12-1 momentum baseline on the Nifty 500, under a strict walk-forward protocol with realistic transaction costs. The result is negative and reported as such: the machine-learning model does not beat the baseline once statistical significance and costs are applied. Even a native learning-to-rank objective that matches the baseline's Sharpe ratio needs 1.8× the portfolio turnover to get there — the added complexity buys no distinguishable benefit.
Evaluation used the Stationary Block Bootstrap and Jobson–Korkie–Memmel tests. The open-source research harness is published alongside the paper.
Evidence from the Nifty 500 · SSRN, 2026 · independent researcher
A second Nifty 500 study tests the institutional convention of sector-neutrality under realistic Indian statutory friction. Across 10 expanding walk-forward folds and 1,260 out-of-sample trading days, capping each sector at three stocks reduced average concentration (HHI 0.169 → 0.084) but cut CAGR from 42.56% to 33.97% and Sharpe from 1.728 to 1.497. The result is statistically significant under both stationary-block bootstrap and Jobson–Korkie–Memmel testing.
The proposed mechanism is factor dilution via rank loss: sector caps force the strategy out of high-conviction momentum leaders in leading industries and into lower-ranked names in lagging sectors. The selected momentum rank deteriorated from 15.5 to 48.2, with Newey–West HAC regression linking that rank loss to subsequent relative underperformance.
Reducing boundary churn and transaction-cost drag in Indian equities (Nifty 500) · SSRN, 2026 · independent researcher
With a buffer band of 60, average monthly liquidations fall from 9.57 to 4.62 names (−51.7 %), annualised cost drag drops from 163 bps to 78 bps, and after-cost CAGR rises from 41.70 % to 45.12 % in the paper's sample (Sharpe 1.728 → 1.849) — a sample-specific result, not a live-trading expectation. The Sharpe gain is not statistically significant (p = 0.167), and the paper says so — the operational case for fewer orders is the contribution.
IRJMETS, January 2024 — collaborative editing architecture; 50+ concurrent users at sub-50 ms sync (implemented as CollabCodex).
Cooperative area coverage with a TurtleBot3 swarm in ROS 2 Jazzy + Gazebo.
- MAPPO / CTDE with a centralized critic for multi-agent credit assignment
- A permutation-invariant GNN policy that scales to any fleet size without retraining — neighbour embeddings pooled through a shared MLP, invalid agents masked
- A Control Barrier Function safety filter solving a QP at 20 Hz to guarantee zero collisions, with nominal-control fallback so a failed solve can't deadlock the robot
- Active Collision Avoidance Safeguard, range-limited consensus map sync, semantic vision, 100 % area coverage in continuous-exploration evaluation
- 25-run, 5-world MAPPO-vs-frontier-heuristic benchmark with a reward ablation under Welch's t-tests — MAPPO reported as a negative result, not buried · archived on Zenodo
Python · ROS 2 Jazzy · Ray RLlib · PyTorch · PettingZoo · Gazebo
Ground-control and Software-in-the-Loop suite for autonomous drone-show fleets, built with an invited contributor team.
- Hungarian-assignment formation morphing between choreography states
- Layered safety loop: pre-flight launch-geometry check (0.5 m tolerance), live proximity rings (2.0 m lateral separation → immediate RTL), a 30 m cylinder geofence triggering emergency landing
- FastAPI + MAVSDK gRPC-to-REST bridge streaming telemetry at 5 Hz over WebSockets; React Three Fiber 3-D previewer;
.skycimport; RTK and latency telemetry - Studio and GCS now live in one monorepo:
tara-core(single source of truth for safety limits),tara-sim(vectorized, 100+ drones), and ataraCLI — one command up, one Ctrl+C down - Georeferenced CesiumJS Earth view plus a 1,000-drone night-sky audience viewport with a live deviation HUD — the show as the ground sees it
- Asymmetric 10 Hz broadcast downlink and 0.5 Hz slotted TDMA uplink; downwash-cone penalty folded into Hungarian assignment; modular scene sequencer and director timeline
- Authenticated command path: Ed25519-signed trajectory uploads, HMAC-SHA256 + anti-replay on the broadcast link, and a geofence envelope computed from the show's own trajectory data
- Operator RBAC, a fleet-lifecycle state machine that 409s illegal commands, pre-arm interlocks, and selective per-drone abort without killing the show
- Fail-closed export gate, a 120 m ceiling enforced at every backend gate, post-rounding re-certification, and every status/audit record labelled with its simulation fidelity tier
Python · PX4 · Gazebo · MAVSDK · FastAPI · React · Three.js · CesiumJS
RoboRashtra 2026 — 1st Runner-Up (₹10,000). A hybrid bot that switches from high-speed deterministic line-following to neural-network inference the moment the maze begins, with a safety supervisor override on top. Spatial distance data → motor commands through hidden layers, running on-device; motor encoders mapped pulse-to-distance so a 90° turn is exactly 90°.
C++ · Teensy 4.1 · Neural networks · N20 encoders
A live, 24/7 momentum strategy on the Nifty 500 with the operational scaffolding to match: HMM regime detection, XGBoost experiments, cross-sectional momentum, a local Ollama LLM layer, Telegram alerting, SEBI 2026 compliance, SQLite WAL persistence, automated backup rotation.
What it's actually taught me is failure modes. A partial data-vendor outage returned NaN prices rather than missing ones; price or entry_price let them through, because float('nan') is truthy in Python. Downstream, abs(nan) > 1.0 evaluated to False, so the drift monitor reported "all clear" on top of a corrupted NAV for two days — a fully green dashboard sitting on bad data. Fixed in three layers (filter at source, explicit math.isnan(), monitor treats NaN as hard failure) plus seven regression tests.
The research arm of this system produced the SSRN paper above.
Python · XGBoost · hmmlearn · stable-baselines3 · Upstox V3 · Ollama · Flask
A limit-order-book matching engine written in Rust — price-time priority, the parts of exchange infrastructure where correctness under concurrency is the whole problem.
Infrastructure intelligence from Sentinel-2: automated measurement of construction progress from multispectral satellite imagery. Train/test separation is enforced in CI, not in good intentions — the pipeline fails the build rather than trusting a developer to remember.
Python · Sentinel-2 · Rasterio · scikit-learn
DEM + two station coordinates → permit-grade aerial ropeway alignment (towers, catenary, safety envelope, DXF, LandXML, Google Earth KML, 3-D twin, BOM, capex) in under 15 seconds on a laptop.
- Validated against 19 real installations across 6 system archetypes on 5 continents — Aiguille du Midi 0.6 %, Zugspitze 2.7 %, Whistler Peak2Peak 7.2 %
- Catenary mechanics validated to 0.00 % against three textbook references (sag / length / tension)
- Eurocode envelope: EN 12929-1 clearance · EN 1991-1-4 wind · ISO 12494 ice · EN 1993-1-1 tower buckling · ANSI B77.1 live load
- Four optimizers benchmarked head-to-head: DEAP GA · pymoo NSGA-II · stable-baselines3 PPO · scikit-learn RSM surrogate
- REST service with JWT auth, RBAC, platform-owner role separation, audit log, immutable project snapshots; 157 pytest cases, CI on every push
- Running in production behind a systemd unit and a Cloudflare tunnel, with Postgres PITR and a verified restore
Python · FastAPI · PostgreSQL · PyVista · Docker · Copernicus GLO-30 DEM
Linear scheduling (LoB / LSM) for heavy civil infrastructure — highways, railways, pipelines, ropeways. Where a Gantt chart shows only time, this plots time against chainage. Phase 15 added a resource-leveling engine, print composer, and DXF/raster backgrounds.
Browser-based simulator for on-demand pod-transit (ATN) networks: synthetic-population OD demand, policy-gradient dispatch, infra cost/ROI modelling against rail and BRT, SUMO/MATSim interop, 481 tests. Era 2 adds vendor-agnostic vehicle profiles, Right-of-Way exclusion zones with cost penalties, KML/KMZ + GeoJSON round-trip, and an Expo SDK 56 rider app.
| Project | What it is | Stack |
|---|---|---|
| AtherCopilot | RAG workspace for real-time PDF & URL analysis — embedding store, streaming responses, Clerk → Firebase token-minting bridge for auth-to-DB identity propagation | Next.js 15 · Gemini 2.5 Flash · Google Genkit · Firebase |
| airgapped-netops-ai | NOC copilot that predicts network degradation and explains itself — zero outbound traffic. Citation-aware RAG, Isolation Forest anomaly detection, SHAP explainability, retraining feedback loop | FastAPI · React · Ollama · ChromaDB |
| Project Morpheus 2026 | Logistics engine + site for a national 24-hour hackathon — 150+ teams, 5 domains, ₹2,00,000+ in prizes. Rule-driven allocation, atomic MongoDB transactions preventing double-booking under concurrent load. Live | Next.js · MongoDB · Three.js · Framer Motion |
| CityConnect | Civic reporting with on-chain EcoCoin rewards — 1000+ req/s with Redis caching | Next.js · Django · Solidity · MongoDB |
| Telhan Sathi | AI oilseed-farming advisory — SIH 2025 National Winner (MoA&FW PS 25273). ML crop-shift models, IoT soil sensing, PWA with real-time farm dashboards | Python · Django · IoT · PWA |
| AgriTrace | Circular-economy platform connecting farmers to recycling agents — AI waste classification + Razorpay traceability ledger | Next.js · Google GenAI · Razorpay · Firebase |
I would rather publish a negative result than a flattering backtest.
My trading research has killed four hypotheses for every two it promoted, and my first SSRN paper reports a model that didn't work. My satellite pipeline enforces train/test separation in CI rather than in good intentions. My drone GCS refuses takeoff over a 0.5 m placement error.
A recent example: mid-project on a helmet-detection model sitting at 368 training images, the obvious move was a 91,000-frame, 29 GB dataset. Instead I retrained on 25/50/75/100 % of what I already had and plotted the curve — 0.49 → 0.62 → 0.70 → 0.72 mAP@50, with the last quarter of data buying +0.02. The model wasn't data-limited. I skipped the download and fixed label quality instead. The technique is page one of any ML course; what's rare is running it before scaling the dataset rather than after.
This is the same instinct across every domain I work in: measure the thing before you buy it, and let the measurement be allowed to say no.
Apr – Oct 2026 · internal enterprise platform team
Built DoppelDash, an internal operational platform (TypeScript, React, Node.js, MongoDB) — role-aware global search, approval chains, audit logging with plain-English summaries, IMAP inbox integration, branded PDF/Excel exports, and a full-application security-hardening pass. 📰 Featured in Maval Maza — state-level press coverage of the selection.
Finalist, Team ID 10144 · official shortlist
Leading GPS-denied autonomous UAV navigation: Fast-LIO SLAM over Livox MID-360, EKF2 sensor fusion with optical flow, ROS 2 + Nav2 + PX4 SITL, Pi 5 / Pixhawk 6c companion-compute split over MAVLink, Coral TPU thermal + RGB fusion for search-and-rescue, and a custom docking station with wireless charging for persistent missions.
Team Algo Sapiens · Dec 2025
Won nationally with Team Algo Sapiens among 10,000+ teams with Telhan Sathi — the first team from SIT Lonavala ever to reach nationals, and to win. Architected the ML crop-shift models and IoT soil-detection hardware, interfaced the hardware to the app live when it failed during the 36-hour finale, and scripted the pitch sequence itself.
Oct 2025 – present · Institution's Innovation Council, SIT Lonavala
Overseeing web, security and blockchain across campus initiatives for 500+ students. Lead Organizer of Project Morpheus 2026 — a national 24-hour hackathon with 150+ teams across 5 domains and ₹2,00,000+ in prizes. Co-architected Kalpataru 2025. Mentored 70+ developers through the CODE2LAUNCH workshop; evaluated 50+ projects on architecture and security.
Robotics & autonomy — ROS 2 (Jazzy) · Nav2 · PX4 SITL · MAVSDK · MAVLink · Gazebo · Isaac Sim · RViz · Pixhawk 6c · Livox MID-360 LiDAR · Fast-LIO · EKF2 · Raspberry Pi 5 · Coral TPU · Teensy 4.1 · motor encoders · PID control
Machine learning — PyTorch · Ray RLlib · PettingZoo · MAPPO/CTDE · graph neural networks · Control Barrier Functions · XGBoost · hmmlearn · stable-baselines3 · scikit-learn · SHAP · TensorFlow · Google Genkit · Gemini 2.5 Flash · Ollama · RAG / ChromaDB
Languages — Python · TypeScript · JavaScript · Rust · C++ · Solidity · Java · SQL · Bash
Backend & data — FastAPI · Node.js · Express · Django · PostgreSQL · PostGIS · MongoDB (transactions) · Redis · SQLite WAL · Prisma · Firebase · WebSockets · gRPC · Celery · Ray
Frontend — Next.js 15 · React · React Three Fiber · Three.js · Electron · Expo / React Native · Tailwind · Framer Motion
Geospatial — Sentinel-2 · Copernicus GLO-30 DEM · SRTM · PostGIS · Rasterio · PyVista · CesiumJS · LandXML · DXF · KML/KMZ · GeoJSON · SUMO / MATSim
Blockchain & Web3 — Solidity · Ethereum · smart contracts · Hyperledger Fabric · IPFS
Infrastructure — Docker · GitHub Actions · AWS · GCP · Vercel · Linux · pytest · Playwright
All 101 public repositories — expand for the full list
| Project | What it is | Stack | Status |
|---|---|---|---|
| 🛰️ Taramandal-GCS | Drone-swarm monorepo — tara-core unified safety limits, tara-sim vectorized simulator, tara CLI, 1,000-drone night-sky viewport with deviation HUD, broadcast/TDMA link layer, point-cloud voxelizer, downwash-cone optimizer, scene sequencer, CesiumJS Earth view, .skyc import; Ed25519-signed trajectories, HMAC-authenticated broadcast, dynamic geofence, 1,000 sim drones in real time; operator RBAC, fleet-lifecycle FSM, pre-arm interlocks, per-drone abort; single-drone V1 mode, fail-closed export gate, fidelity-tier labelling |
Python · PX4 · FastAPI · React · Three.js | Public · V1 roadmap · contributor team |
💰 Autonomous-Ropeway-Alignment (RopeIQ) private |
Eurocode-validated ropeway alignment optimizer — 19 validated real installations, four optimizers, DXF/LandXML/KML export, permit dossier with CAPEX + span summary, Postgres PITR with verified restore | Python · FastAPI · PostgreSQL · PyVista | In production · systemd + Cloudflare tunnel |
| 🛡️ netguard-ml | Leakage-free ML intrusion detection on CIC-IDS2017 — honest evaluation protocol, zero-day evaluation across both corpora, real pcap / live-capture path via NFStream, REPRODUCE.md for a clean-clone rebuild |
Python · scikit-learn · PyTorch · NFStream | Public |
| 🚚 fleetcommand-platform | FleetCommand V2 — Flutter mobile + Spring Boot backend, phase roadmap and a go/pivot review template gating pilot success | Flutter · Spring Boot · C++ | Scaffold · V2 |
| 🚦 traffic-surveillance-system | Adaptive spatio-temporal traffic surveillance, now public — vehicle-box-first tracking with helmet checks gated to two-wheelers, multi-frame wheel-count classification, ONNX export | Python · DeepSORT · OSNet · ONNX | Public · paper in review |
| 🤖 MARS | Multi-agent TurtleBot3 swarm — MAPPO/CTDE, permutation-invariant GNN policy, 20 Hz CBF safety QP, 100 % coverage; 5-world benchmark with MAPPO reported as a negative result, failure-to-reproduce companion report, Zenodo DOI | Python · ROS 2 · Ray RLlib · PyTorch | v1.0.0 · Zenodo DOI |
| 🚛 SIH26137 | FleetOps command-center prototype for a Smart India Hackathon 2026 problem statement — mobile-responsive slide-over nav, touch map, sheet modals, command palette | TypeScript | Prototype |
| 🌾 kisan-micro-erp | Offline-first micro-ERP for small farm businesses — double-entry engine, deterministic transaction parser, SQLite-WASM/IndexedDB, TB/P&L/balance sheet, Ed25519-signed backups | TypeScript · PWA · SQLite WASM | MVP complete |
| 🤖 Bot | Teensy 4.1 neural maze solver + line follower — RoboRashtra 2026 1st Runner-Up | C++ · Teensy 4.1 · Neural networks | Competition winner |
| 🌐 airgapped-netops-ai | Fully offline NOC copilot — citation-aware RAG, Isolation Forest anomaly detection, SHAP explainability; 5-job CI with coverage gate & secret scanning | FastAPI · React · Ollama · ChromaDB | Tracks 1–3 · CI hardened |
| ⏱️ ChronoTrack | Privacy-respecting workforce telemetry & project-ROI platform — Electron agent + Python AI daemon + web dashboard, GDPR/DPDP-compliant, 2FA with TOTP & passkeys | Electron · Python · React · PostgreSQL | v3.2.5 |
| 🤖 AtherCopilot | RAG AI workspace — multi-document context, Clerk → Firebase auth bridge | Next.js 15 · Genkit · Gemini · Firebase | Live |
| 🎯 Project-Morpheus | National 24-hour hackathon platform — 150+ teams, atomic allocation engine | Next.js · MongoDB · Three.js | Live |
| 🌱 AgriTrace | Agricultural waste traceability — AI classification + Razorpay ledger | Next.js · Google GenAI · Firebase | Active |
| 🎮 3d-chess | 3-D multiplayer chess — Stockfish AI, Supabase realtime, glassmorphic UI | Next.js · React Three Fiber · Supabase | Live |
| 🚁 Drone | Scrollytelling drone-workshop site — 40-frame scroll sequence, isometric LiDAR swarm sim on Canvas | Next.js 14 · TypeScript · Canvas API | Live |
| 🚁 Fly | ASCEND — aerial intelligence landing page (autonomous flight, precision data capture) | HTML · JS · 3D | Live |
| ❄️ SmartCooler-IoT | ESP32 cooling controller — relay + DHT22 + DS18B20, humidity auto-mode, NVS persistence, web dashboard | ESP32 · C++ · Firebase | Deployed |
| 😴 Drowsiness-Detection-System | Real-time eye-aspect-ratio drowsiness detector for driver safety | MediaPipe · Node.js · OpenCV | Built |
| 📅 365DaysOfCodeChallenge2k25 | A year-long daily coding challenge, in public — CodeQL-clean, dependency CVEs patched via overrides | Python · Web | Day 139 |
| 📡 Lidar · thermal · UWB | Sensor-integration research modules for the ASCEND / IRoC drone stack | Python | Research |
🛰️ iroc-ascend-nav private |
IRoC drone nav stack — simulation-first on a 1-year pure-sim roadmap; pure-Python 3-D occupancy mapping + 3-D A* planning, Three.js dashboard, Mars heightmap terrain | Python · ROS 2 · NumPy · Three.js | Simulation-first |
📊 AlgoTrade private |
Systematic NSE momentum system — HMM regime, XGBoost research, Ollama LLM layer, Telegram ops, SEBI 2026 compliance; isolated real-money live pilot track with fail-closed kill switch, capital ceiling, order idempotency and emergency liquidation; date-effective F&O instrument, cost and margin models with a falsification-first research track; source of the SSRN paper | Python · Upstox V3 · XGBoost · Flask | Runs 24/7 · live pilot · F&O research |
🛰️ terra-verify private |
Construction progress measured from Sentinel-2 imagery; train/test separation enforced in CI. Formal backtest scorecard over six real sites incl. sourced negative cases, PIF land-cover limitation documented rather than papered over, AOI-draw UI with SSE job status | Python · Sentinel-2 · FastAPI · Next.js | Backtest scorecard · v0.3.0 |
🚉 pod-transit private |
PRT/ATN digital-twin simulator — OD demand, RL dispatch, cost/ROI vs rail & BRT, SUMO/MATSim interop, 481 tests | React · TypeScript · Three.js · Expo | Era 2 |
🛣️ linear-scheduler-app private |
Line-of-Balance scheduler for highways, railways & pipelines — resource leveling, print composer, DXF backgrounds | Next.js · PostgreSQL · Prisma | Phase 15 |
🛠️ GripGroup private |
Generative geospatial design for linear infrastructure — SRTM/PostGIS, multi-objective optimizer, LandXML export, Ray + Celery, credit-ledger billing | Python · React · PostGIS · Ray | Phase 9 |
🌾 KrishiPancham-AI private |
Agricultural AI companion — Panchang almanac + weather + NDVI crop health + 32 govt schemes; custom-trained crop-disease classifier at 97.4 % val accuracy hybridised with Gemini; multilingual, voice I/O, offline-first | FastAPI · Expo · PyTorch · Postgres | Play Store readiness |
📱 Potnis (ExpenseTracker) private |
Production React Native app — bank SMS parser for 10+ Indian banks, OTA updates, 4-layer dedup sync, biometric app lock, OCR evidence-gate benchmark harness | React Native · TypeScript · MongoDB Atlas | Play Store prep |
🏙️ city-companion private |
Privacy-first mobility telemetry + AI survey generation for traffic analytics | TypeScript · Expo · Express · Prisma | Phase 4 |
🛣️ nextgen-highspeed-toll private |
Zero-hardware, fully simulated highway toll digital twin | React · MongoDB · RabbitMQ · ROS 2 | MVP |
💇 queuepay private |
QR-scan-to-pay salon queue booking — Next.js web + Flutter Android on Supabase, RLS-first data model, labelled good-first-issue backlog, Dependabot + CI from day one; first outside contribution (architecture flow diagram) merged | TypeScript · Flutter · Supabase | Phase 1 MVP |
🧾 rentclear private |
Rent payment that stays TDS-compliant — s.194-IB / Form 26QC / Form 16C, UPI rail, PAN verification behind an Edge Function, citations re-grounded on the Income-tax Act 2025 | Flutter · Supabase · Razorpay | Phase 1 |
🚕 RideRadar private |
Cab-fare aggregator across Uber/Ola/Rapido/Namma Yatri — real ONDC Beckn subscriber layer, TRV10 ride-hailing and TRV11 transit-to-QR-ticket flows, luggage/payload filter | Python · ONDC Beckn · Expo · Appium | ONDC pre-prod |
⛽ fuel-margdarshak private |
Fuel-station finder with Android Auto, route search, territory game and a B2B partner analytics dashboard — offline-first SQLite with observation-time conflict resolution, favourites/My Places, Devanagari keyword fallback | Flutter · Supabase · Android Auto | Play Store prep |
| 📷 esp32cam-01…06 | Six teaching-grade ESP32-CAM builds — live stream, SD-card timestamping, photo booth, timelapse, PIR security, Telegram alerts | C++ · ESP32-CAM | Published |
🧠 Harsh OS private |
Local-first 4-agent cognitive layer for LinkedIn — Tampermonkey ↔ Ollama, local vector DB, zero data leaves the machine | Vanilla JS · Ollama | v1 |
💼 DoppelDash private |
Internal operational platform built during the Doppelmayr internship — now multi-tenant with full orgId route scoping, organization self-signup, and mobile Bearer-token auth with refresh-token rotation; native Flutter Android client with TOTP 2FA, Play-subscription entitlements and a 442-test CI gate; web CI at 1,071 tests through staged launch verification |
TypeScript · React · Node.js · MongoDB · Flutter | Play Store prep |
🛰️ cubesat-1u private |
1U CubeSat engineering programme — systems, flight software, electronics, mechanical and ground-segment tracks, first sprint on a telemetry vertical slice | Shell · Embedded | Sprint 1 |
| Project | Contribution |
|---|---|
| validatorjs/validator.js | #2713 — fixes the uz-UZ mobile-phone regex (adds missing carrier prefixes); open, awaiting review |
| Textualize/rich | #4129 — docs: removes a dead link from the README (fixes #4123); open |
- Smart India Hackathon 2025 — National Winner · first team from SIT Lonavala ever to win nationally · certificate
- ISRO Robotics Challenge (IRoC-U 2026) — Finalist, Automation & Navigation Lead, Team ID 10144 · official listing
- RoboRashtra 2026 — 1st Runner-Up, Chakravyuh maze & line solving (₹10,000)
- Singapore–India Hackathon 2026 — selected for the Grand Finale at Singapore Management University (AICTE & MoE Innovation Cell), 15–17 Nov 2026
- AICTE & MIC — selected for Phase II of the IDE Bootcamp, Ministry of Education · certificate
- Bentley Systems Illuminate 2026 — digital twin / MicroStation / OpenFlows summit participant
- Press — Maval Maza, state-level coverage
- PCU Ideathon 2.0 — Top 40 of 500+ teams
Certifications — Certified Ethical Hacker Bootcamp (ZTM) · Advanced DSA (Codedamn) · Computer Networks (Udemy) · Web Developer Bootcamp 2025 (Udemy) · Data Science 101 (IBM)
B.Tech, Computer Science — Sinhgad Institute of Technology (SIT), Lonavala · Sep 2024 – Aug 2027 CGPA 7.18 · Sem 5 SGPA 8.05 · O grade in Laboratory Practice I, CN & Security Lab, and Seminar & Technical Communication · Blockchain Honours: O (practical), A (theory) · A+ in Distributed Systems Coursework: Data Structures & Algorithms, Operating Systems, DBMS, Theory of Computation, Distributed Systems, Computer Networks
Diploma in Computer Engineering (First Class with Distinction) — SPM Polytechnic, Solapur · 2021–2024 · General Secretary ICSE Class 10 — Presentation Convent School, Nerul · 89 %
Targeting GATE 2027, top-500 rank — Distributed Systems, DBMS, Theory of Computation, Computer Networks.
- dev.to/harshpandhe — 18 posts on robotics, system design, GitHub workflows, and Blender for autonomous systems
- LinkedIn — engineering write-ups: drone-swarm safety architecture, the
float('nan')truthiness bug that corrupted a trading NAV, learning-curve analysis before scaling a dataset, and the SSRN paper announcement - CODE2LAUNCH — workshop leader, 70+ attendees (React, Node.js, JavaScript)
- Infotsav '25 — speaker (CityConnect)
- IRoC-U 2026 Finals — clear the final round with ISRO
- Get the negative-result paper onto arXiv q-fin.PM (seeking endorsement)
- RopeIQ — industry outreach and enterprise validation on the deployed service
- ChronoTrack — enterprise pilot deployment
- pod-transit Era 2 — finish the fleet / ROW / geo-export track
- terra-verify — publish the validation methodology
- GATE 2027 — top-500 rank
- Speak at a robotics or edge-AI conference
- Taramandal V1 — single-drone bench, supervised hover/landing acceptance gate, and an integrated proof demonstrator
- DoppelDash Android — first Play Store release of the Flutter client
- cubesat-1u — close Sprint 1 with an end-to-end telemetry slice (sensor → flight software → ground dashboard)
- Taramandal Tier 1 — music-locked show timing and a hardware-in-the-loop rehearsal beyond SITL
- airgapped-netops-ai — cut a tagged release through the new CI pipeline
- traffic-surveillance-system — land both IJSRSET paper submissions
- Potnis and fuel-margdarshak — first Play Store releases
- netguard-ml — get the live-capture path validated on traffic I didn't generate myself
- AlgoTrade live pilot — run the real-money track to a full reporting window and publish the paper-vs-live slippage honestly
- Taramandal Studio — Web Audio music visualizer and multi-track LED choreography editor (the last open
difficulty:coreissue) - queuepay — close out the Phase 1 MVP backlog with outside contributors
- MARS — submit the swarm-coverage paper (and its negative-result companion) as an arXiv preprint
- SIH 2026 — take the FleetOps command-center prototype from mock to working pilot
- AlgoTrade F&O — take OPT-FLOW-001 from HOLD to a decisive promote-or-kill under the new cost and margin models
- Taramandal Studio AI — local Ollama show copilot, image-to-formation and music-aware choreography (epic open)
- Taramandal V1 end-to-end — a Studio-built show flown in PX4 SITL + Gazebo, plus the arm / ~1 m hover / slow-land single-drone acceptance run
- Taramandal validation at scale — bring
/api/validateat 500–1,000 drones down from minutes to interactive - DoppelDash web launch — execute the production cutover plan after staging sign-off
- ✅ Taramandal goes public — repo opened, with operator RBAC, fleet-lifecycle FSM, pre-arm interlocks and per-drone abort merged
- ✅ kisan-micro-erp MVP — offline-first double-entry ERP finished and QA'd on a physical Android device
- ✅ MARS v1.0.0 — research artifact finalized, companion papers rebuilt on archived data
- ✅ Taramandal link security — Ed25519-signed trajectory uploads, HMAC-authenticated broadcast with anti-replay, 1,000 sim drones in real time
- ✅ MARS on Zenodo — citable DOI, CITATION.cff, machine-portable reproducibility bundle
- ✅ Taramandal at scale — 1,000-drone synthetic telemetry load test, cKDTree spatial benchmark and WebGL night-sky viewport all landed
- ✅ RopeIQ in production — systemd unit, Cloudflare tunnel, Postgres PITR with a verified restore
- ✅ Taramandal monorepo — Studio + GCS unified on
tara-core, opened to outside contributors (first external PR merged) - ✅ IRoC arena — all 15 rule-book features built, N=30 validation campaign at 30/30
- ✅ Taramandal Tier 0 — collision-free formation morphing, verified end-to-end Studio → GCS
- ✅ Taramandal-GCS Phase 3 — RViz/TF integration live-verified on SITL
- ✅ Publish quantitative research to SSRN
- ✅ Ship the Taramandal studio → GCS pipeline
- ✅ MARS Phase 2 — CBF safety filter, collision avoidance, 100 % coverage
- ✅ ChronoTrack v3.2.5 — 2FA, team-lead dashboards, auto-update installer
- ✅ Open-source the maze-solver neural network firmware (Bot)
Open to: research collaborations · open-source work · robotics, autonomy and quantitative engineering roles. Particularly interested in: autonomous navigation · multi-agent systems · geospatial infrastructure · quantitative research · distributed systems.
📍 Navi Mumbai, Maharashtra, India · 🌐 harshpandhe.com · 📧 harshpandhehome@gmail.com
Keywords: autonomous navigation engineer · multi-agent reinforcement learning · ROS 2 developer · PX4 drone software · LiDAR SLAM · Control Barrier Functions · quantitative research India · Nifty 500 momentum · Sentinel-2 remote sensing · ropeway alignment optimization · Harsh Pandhe



