Open to opportunities · Pune, India

Shubham Kulkarni — AI Engineer

I build computer vision that watches real highways.

Software Engineer (AI/ML) at Arya Omnitalk (Arvind Group). I design the detection, tracking and incident pipelines behind live highway traffic management — VIDES, ATMS and TMCS — processing 50,000+ events a day at 95% model accuracy.

Download résumé See the flagship work

now> AI Engineer @ Arya Omnitalk

CAM-07 · PRESET 02
--:--:--
SIMULATION
0 tracks · p95 41 ms
Event streamSSE · live
  1. SYSTEM7 models loaded · 14 detectors armed
Events / day
50,000+
Model accuracy
95%
Inference latency
<100ms
Forensic queries / day
150K+
Faster retrieval
30×
System uptime
99.9%

01 About

Research is easy to demo. I care about what survives production.

Portrait of Shubham Kulkarni
fig. 01 — Pune, Maharashtra

I'm an AI Engineer at Arya Omnitalk Wireless Solutions (part of the Arvind Group), where I design and deploy computer vision and generative AI systems for India's live highway infrastructure.

My work sits where models meet messy reality: PTZ cameras that move, headlights at 2 a.m., riders who look like pedestrians, cameras that drop off the network. My systems handle 50,000+ events daily at 95% accuracy with inference under 100 ms, and I engineered 30× faster vector retrieval with FAISS IVF indexing to serve 150,000+ forensic vehicle queries a day.

Before that I was a National Champion (AIR 1) at Smart India Hackathon 2023 as team lead, interned at IBM and MathWorks, and shipped an alumni portal my college still runs. B.E. in Electronics & Telecommunication, SPPU, 2025.

Current
Software Engineer — AI/ML, Arya Omnitalk (Arvind Group)
Focus
Computer vision · Multi-object tracking · RAG · Edge AI
Education
B.E. E&TC, SPPU '25
Recognition
AIR 1 — Smart India Hackathon 2023
  • Computer Vision
  • Generative AI
  • RAG Pipelines
  • Edge Deployment
  • MLOps
  • Document AI
  • LLMs & Agents
  • Full Stack

02 Experience

From internships at IBM & MathWorks to production AI at scale.

  1. Aug 2025 — Present

    Software Engineer — AI/ML

    Arya Omnitalk Wireless Solutions Pvt. Ltd. (Arvind Group) · Pune · Full-time, on-site

    • Core engineer on VIDES, an edge ATCC & enforcement network running on NVIDIA Jetson Orin at highway gantries: 14 detection types — ANPR, helmet, triple riding, seatbelt, speed, wrong-way, stopped vehicle, accident, pedestrian, animal, pothole — across 50,000+ events/day at 95% accuracy. Case study ↓
    • Engineered TMCS, the incident-detection layer for NHAI ATMS-aligned corridors: dual-model detection, occlusion-resilient tracking, a 7-rule incident engine and tamper-evident evidence. Case study ↓
    • Ran 7 models (6 TensorRT FP16 engines + PP-OCRv4) side by side on one Jetson with GPU-direct GStreamer decode, holding frame processing under 40 ms across up to 6 concurrent 1080p streams.
    • Built the ANPR path — plate localizer, PaddleOCR, blur penalties and 2-read consensus — plus parallel helmet and seatbelt pipelines fused by a per-vehicle decision finalizer.
    • Integrated NVIDIA Jetson edge nodes with backend microservices over async REST and an SSE → Socket.IO relay; multi-threading and batching cut end-to-end latency by 30%, with alerts reaching operators in under 150 ms.
    • Deployed across AWS/Azure and edge with CI/CD at 99.9% uptime, working across hardware, backend and operations teams.
    • Production AI
    • Computer Vision
    • Edge
    • MLOps
  2. Oct — Dec 2024

    Data Analytics & GenAI Trainee

    AICTE · VOIS for Tech · Vodafone Idea Foundation · Remote

    • Built automated multi-source ETL pipelines in Python & Pandas ingesting IoT soil moisture, ambient temperature, and satellite weather telemetry across agricultural monitoring nodes.
    • Applied open-source Large Language Models (LLMs) and LangChain orchestration to synthesize multi-variate telemetry into automated natural-language agronomic diagnostics and crop stress advisories.
    • Modeled statistical trend projections and time-series forecasting for precision irrigation optimization, benchmarked against FAO-56 Penman-Monteith evapotranspiration formulas.
    • Engineered data validation and outlier-cleansing routines on multi-gigabyte datasets, curtailing downstream analytical pipeline anomalies by 40%.
    • GenAI & LLMs
    • Data Engineering
    • ETL Pipelines
    • LangChain
    • Agritech Analytics
  3. Aug 2023 — Aug 2024

    Campus Ambassador & Technical Lead

    GeeksforGeeks · Ahilyanagar · Campus Leadership

    • Spearheaded the official GeeksforGeeks collegiate chapter at Dr. VVP COE, organizing 5+ hackathons, algorithm sprint contests, and technical bootcamps.
    • Mentored and coached over 200+ undergraduate engineering students in Data Structures & Algorithms (graph theory, dynamic programming), Python optimization, and interview problem-solving.
    • Conducted hands-on workshops on Machine Learning and Computer Vision fundamentals, elevating active campus developer contest engagement by 150%.
    • Liaised directly with GeeksforGeeks regional leadership to secure event licensing, contest prize pools, and verified developer training material for students.
    • Technical Mentorship
    • DSA & Algorithms
    • Hackathons
    • Community Leadership
    • Competitive Coding
  4. Jun — Aug 2023

    Artificial Intelligence Intern

    IBM India Pvt. Ltd. · Bangalore / Remote · Technical Internship

    • Engineered and fine-tuned transformer-based NLP text classification models (BERT, DistilBERT, RoBERTa) achieving 90% classification accuracy on psychological sentiment and distress benchmarks.
    • Architected the Mental Health AI Tracker — an on-device behavioral telemetry system using Exponentially Weighted Moving Average (EWMA) lexical valence drift for early cognitive burnout detection. Case study →
    • Implemented Laplace Differential Privacy (ε = 0.5) with local perturbation vectors to mathematically protect sensitive clinical user inputs against reconstruction attacks.
    • Quantized transformer weights to INT8 via ONNX Runtime, achieving sub-15 ms on-device inference latency with zero raw text telemetry leaving the client runtime.
    • NLP & Transformers
    • BERT / RoBERTa
    • Differential Privacy
    • ONNX Quantization
    • Sentiment Analysis
  5. May — Sep 2023

    Computer Vision & Deep Learning Intern

    MathWorks · Ahilyanagar / Pune · Technical Internship

    • Engineered deep learning image processing workflows and convolutional neural network (CNN) feature extractors in MATLAB for automated visual inspection and object categorization.
    • Developed image preprocessing and enhancement pipelines using CLAHE, bilateral edge-preserving filters, and morphological transforms for degraded and low-contrast image captures.
    • Built automated image augmentation pipelines (affine warping, rotations, chromatic perturbation) expanding training data variety and improving model generalization by 14%.
    • Profiled and benchmarked CNN layer activations, parameter footprints, and inference latency using MATLAB Deep Learning Toolbox and GPU Coder for embedded target verification.
    • Computer Vision
    • MATLAB & Simulink
    • CNN Architectures
    • Image Processing
    • Deep Learning Toolbox
  6. 2021 — 2025

    B.E. in Electronics & Telecommunication Engineering

    Dr. Vithalrao Vikhe Patil College of Engineering · Savitribai Phule Pune University (SPPU)

    • AIR 1 National Winner — Smart India Hackathon 2023: Led the collegiate engineering team to All India Rank 1 out of 50,000+ applicant teams for an AI-powered smart bin sorting and route optimization system. Inside story →
    • Official University Alumni Portal: Designed and deployed the college's production Alumni Management Portal (Django 5, PostgreSQL, Redis, Socket.IO, 50+ REST endpoints), officially adopted and actively utilized across 4 departments and hundreds of alumni. Architecture deep dive →
    • Precision AgTech Innovation: Co-developed JalTantra, an IoT + ML precision irrigation solution saving 30% water (University Rank 2, SPPU Startup Olympiad 2025). Project overview →
    • Core Committee Secretary: Served on the apex student governance committee, managing budget planning, logistics, and student operations for national technical symposia and collegiate events.
    • SPPU
    • SIH 2023 AIR 1
    • Full-Stack Architecture
    • IoT & Embedded
    • Student Governance

03 Flagship case study · Arya Omnitalk

VIDES — an edge enforcement network that audits every vehicle, not just the violators.

VIDES (Vehicle Identification & Detection System) is the ATCC and traffic-enforcement platform for expressway gantries and toll corridors. Each gantry runs its own NVIDIA Jetson Orin node: camera streams are decoded straight into GPU memory, every vehicle is detected, tracked, classified and plate-read, and violations and incidents are confirmed at the roadside. The control room receives structured, evidence-backed events — not raw video.

Most enforcement systems only wake up for violations. VIDES is built on complete verification: every clean passage is logged too, with class, speed, plate and evidence, so the corridor doubles as a fraud-resistant toll-auditing network.

Role
Core engineer — edge pipeline, tracking, ANPR, violation & incident engines, real-time relay
Hardware
NVIDIA Jetson Orin NX / Nano · TensorRT FP16
Per node
Up to 6 concurrent 1080p RTSP streams
Status
Production system · proprietary

Detection types

14

violations, incidents & audit

Models on one node

7

6 TensorRT FP16 engines + PP-OCRv4

Frame processing

<40ms

on-device, per stream

Edge → operator

<150ms

SSE → Socket.IO alert

Events per day

50K+

at 95% model accuracy

Vehicle classes

7

car · truck · bus · mini bus · auto · bike · tractor

Plate lock

2-read

OCR consensus before a plate is final

Decision timeout

6s

finalizer never waits on a slow model

Architecture one edge node per gantry

Two camera roles share one perception core, then split into enforcement and incident paths.

  1. 01RTSP → GPU decodeGStreamer nvv4l2decoder · latest-3-frame buffer
  2. 02Vehicle detectionYOLOv8 · TensorRT FP16 · 7 classes
  3. 03TrackingByteTrack · trajectories · speed gates
  4. 04Early rejectionROI · lane · direction · size · track quality

ANPR / enforcement cameras

  • ANPRplate localizer → PP-OCRv4 · 4 workers · 2-read consensus
  • Helmet & triple ridingtwo-wheelers only
  • Seatbeltwindshield region · 4+ wheelers
Detection finalizerfuses parallel results per vehicle → violation or clean audited passage

Overview / incident cameras

  • Trajectory ruleswrong-way · stopped · queue · accident
  • RT-DETRpedestrians & animals · every 3rd frame
  • Pothole scanhourly · daylight only
Incident managerevent fusion · 5-minute dedup cooldown
Async DB writer → PostgreSQLtracking threads never block on SQL
SSE → Socket.IO → operator dashboardverify / discard with evidence, <150 ms

14 detection types and the rule behind each

Critical hazards

  • Wrong-way drivingmoves against the lane vector for 5 consecutive frames → critical alarm
  • Stopped vehicle<10 px movement for 40 s at ≤2 km/h inside a lane
  • Vehicle accident≥3 vehicles stopped within a tight spatial cluster

Environmental & lane hazards

  • Pedestrian on highwayRT-DETR, 0.65 confidence held for 8 frames; safe zones excluded
  • Animal on roaddog, cow, buffalo, sheep — same confirmation gate
  • Wrong-lane usagevehicle class checked against the lane's allowed classes
  • Pothole / road damagescheduled daylight scans; evidence for maintenance crews

Flow & audit

  • Audited passageevery clean vehicle logged with class, plate and evidence
  • Overspeedingperspective-corrected gate timing vs the lane limit; impossible speeds discarded
  • Slow-moving vehiclebelow the lane's minimum speed in fast lanes
  • Congestion≥5 tracks averaging under 15 km/h

Occupant compliance

  • Helmet missingno-helmet ratio over a 5-second window
  • Triple riding≥3 riders tied to one two-wheeler track
  • Seatbelt missing≥2 driver votes; configurable rule for occluded windshields

Hard problems, and how I solved them

A

Zero-copy ingest, no encoder

Streams decode straight into GPU memory via nvv4l2decoder. The Jetson Orin Nano has no hardware encoder, so a GStreamer tee forwards the original H.264 to live preview without re-encoding.

B

Capture never waits

Each camera keeps only its latest 3 frames. If inference hiccups, stale frames are dropped rather than queued, so latency stays bounded instead of drifting behind real time.

C

Reject early, compute late

ROI, lane, direction, size and track-quality filters drop irrelevant tracks before any expensive model runs — only vehicles that matter reach OCR, helmet or seatbelt inference.

D

Plates you can trust

CLAHE and sharpening before OCR, a Laplacian blur penalty, position-aware O/0 and I/1 fixes, and a plate only locks after 2 matching reads.

E

Fan-out, fan-in decisions

ANPR, helmet and seatbelt run in parallel worker pools on bounded queues. A finalizer fuses their results per vehicle with a 6 s timeout, so one slow model never stalls a record.

F

Seven models, one small GPU

TensorRT engines share a GPU lock so contexts never collide on 8–16 GB Jetsons. RT-DETR runs every 3rd frame and pothole scans hourly, keeping the whole stack inside the power and memory budget.

G

The hot path never touches SQL

All writes go through an async DB writer. The dashboard relay reconnects with exponential backoff and keeps a rolling per-camera cache, so operator screens load instantly without hitting the database.

H

Humans confirm, machines log

Every violation reaches an operator with evidence to verify or discard — discards are retained for audit. Admin, operator and viewer roles gate who can change ROIs, cameras and thresholds.

Why the edge

Central / cloud ATCCVIDES edge node
Latencynetwork round-trip before any decisiondecided at the gantry in <40 ms
Bandwidthcontinuous raw-video uplinkstructured events + evidence crops
Privacyraw footage leaves the siteframes processed on-site; only evidence kept

Stack

  • Python
  • Flask
  • GStreamer
  • OpenCV
  • YOLOv8
  • RT-DETR
  • ByteTrack
  • PaddleOCR PP-OCRv4
  • TensorRT FP16
  • CUDA
  • NVIDIA Jetson Orin
  • PostgreSQL
  • SSE
  • Socket.IO
  • MediaMTX
  • systemd

VIDES is a proprietary Arya Omnitalk system. Source code, deployment sites and internal documentation are confidential; the architecture and figures above are shared at a level suitable for public discussion.

04 Flagship case study · Arya Omnitalk

TMCS — turning every highway camera into an automated incident detector.

TMCS (Traffic Monitoring Camera System) converts live IP/PTZ camera streams into explainable, evidence-backed incident events, built against NHAI's ATMS Standards & Specifications (Oct 2023), which place monitoring cameras roughly every kilometre of national highway. It is the detection layer beneath the highway ATMS: it does the continuous watching and hands the control room a structured, versioned, audit-traceable event — not a raw video feed to be reinterpreted later, under pressure.

A control room can't scale headcount with camera count. TMCS closes that gap at the camera, across multi-vendor ONVIF hardware, day and night — without replacing the cameras already installed.

Read the solution overview PDF · 14 pages
Role
Core engineer — vision pipeline, tracking, rule engine, streaming, PTZ, evidence & RBAC
Domain
Highway ATMS · NHAI ATMS Standards, circular 11.53/2023
Scale target
25–30 FPS per camera · multi-camera GPU scheduling
Status
Production system · proprietary
Delivery
8 phases, each commissioned and signed off — pilot camera to full corridor

Camera feed → confirmed incident. Nothing becomes an alert until every stage agrees.

  1. 01Camera · ONVIFmulti-vendor PTZ
  2. 02RTSP ingest25–30 FPS, auto-reconnect
  3. 03Preset scene gateROIs per PTZ preset
  4. 04DetectionYOLOv8 + RT-DETR · TensorRT
  5. 05TrackingByteTrack · dt-aware Kalman
  6. 06Rule engine7 temporal-spatial rules
  7. 07Confirmed eventdedup · lifecycle · audit
  8. 08Evidence & hand-offWORM + SHA-256 · SSE → ATMS

Benchmarks internal engineering measurements

Pipeline latency (tracker + rules, excl. model)

10 objects1.5 ms
40 objects5.3 ms
80 objects10.3 ms

p50 per frame. Scales roughly linearly with scene density.

Inference target

≤45ms p95

Dual-model, TensorRT FP16/INT8

Preset sync

2,500 → <10ms

−99.6% via one batched SQL upsert

Dead-camera rejection

0.1ms

vs. 10–30 s socket timeouts

Shutdown

<0.5s

from 10–30 s thread hangs

Analytics SQL load

−90%+

60 s in-memory TTL aggregation

Automated tests

298+

unit, integration, fault-injection

Track retention

60s

pedestrians & animals behind trucks

False pedestrian alarms

−95%+

rider-aware occupant tagging

Hard problems, and how I solved them

A

Dual-model detection

YOLOv8 for vehicles plus RT-DETR for pedestrians, animals and debris, merged with class-agnostic NMS — each model does what it's best at, on one GPU-resident engine cycling across cameras.

B

Tracking through occlusion

ByteTrack with a dt-aware Kalman filter (xyah). Category-aware retention keeps a pedestrian's or animal's identity for up to 60 s while a truck passes in front.

C

Riders aren't pedestrians

Spatial occupant tagging geometrically links people to two-wheelers and auto-rickshaws, cutting false pedestrian alarms by over 95%.

D

Night-mode glare guard

EMA frame-luma tracking and an HSV saturation filter reject headlight glare and wet-road reflections. At night the bar for new tracks rises by +0.15, while confirmed tracks keep normal thresholds. No wall-clock assumptions — tunnels are dark at noon.

E

PTZ-safe analytics

Every PTZ move or preset recall clears tracks and pauses analytics until the camera stabilizes, then loads that preset's own ROIs. A moving camera never fires a false wrong-way against the wrong geometry.

F

Fail-fast device layer

An ONVIF supervisor with a 20 s circuit breaker keeps dead cameras from blocking async workers. Re-entrant locks and a watchdog guarantee deterministic shutdown; Pelco-D presets fall back to AbsoluteMove.

G

Court-grade evidence

Write-once evidence with a SHA-256 digest, never edited by later processing. Boxes, track ID and class are rendered as an overlay on the untouched image, and every event records the exact model, rule and scene-config versions that fired it.

H

Enterprise integration & access

Shares the master VMS/ATCC database through a synchronized view, so it coexists with legacy systems with zero downtime. RBAC with 36 permissions across 6 domains and 5 role tiers, cached with instant invalidation and full audit logging.

A layered platform each layer owns one job

  1. DeviceONVIF discovery, capability negotiation, credentials, PTZ control
  2. MediaRTSP ingest, decode, timestamping, buffering, auto-reconnect, stream health
  3. ScenePTZ preset state, stabilization, zones, lanes, direction, calibration
  4. AIdetection, multi-object tracking, model lifecycle, GPU execution
  5. Rulestemporal & spatial checks, thresholds, candidate confirmation
  6. Eventidentity, lifecycle, dedup, evidence orchestration, audit trace
  7. Integrationhand-off to ATMS, VMS message signs, patrol & emergency dispatch

Operational trust how changes reach a live highway

  • Config lifecyclepresets, zones, rules and models go draft → validate → approve → publish; history is never overwritten
  • Proven off-road firstsimulator, recorded footage and shadow mode — no physical cameras needed to test a rule
  • Secrets stay secretcamera credentials and API tokens never live in code or logs
  • Honest state"command accepted" is never shown as "action completed"; features degrade gracefully
  • Auto PTZ returnafter manual inspection, cameras return to their surveillance preset on a timeout

7 incident rules · 11 object classes

  • Stopped vehicle
  • Wrong-way driving
  • Pedestrian on highway
  • Animal on road
  • Debris
  • Queue / congestion
  • Parked vehicle

Car · Motorcycle · Truck · Bus · Mini bus · Auto-rickshaw · Tractor · Bicycle · Pedestrian · Animal · Road debris

Stack

  • Python 3.11
  • FastAPI
  • asyncpg
  • PostgreSQL
  • PyTorch
  • YOLOv8
  • RT-DETR
  • TensorRT
  • ByteTrack
  • OpenCV
  • ONVIF
  • RTSP
  • MediaMTX
  • MJPEG
  • SSE
  • HTML5 Canvas

TMCS is a proprietary Arya Omnitalk system. The linked overview is Arya Omnitalk's client-facing solution document; source code and client deployment details stay confidential, and figures above are internal engineering benchmarks and acceptance targets.

05 Selected work

Production systems, national wins & shipped platforms.

01Flagship · Production

VIDES — Edge Traffic Enforcement Network

ATCC and enforcement on NVIDIA Jetson Orin at every gantry: detection, tracking, ANPR, helmet, triple riding, seatbelt, speed, wrong-way, accident, pedestrian, animal and pothole detection — confirmed at the roadside and streamed to operators in under 150 ms.

Detection types
14
Models per node
7
Frame budget
<40 ms
Events/day
50K+
  • YOLOv8
  • RT-DETR
  • ByteTrack
  • PP-OCRv4
  • TensorRT
  • GStreamer
  • Jetson Orin
Case study ArticleProprietary
02Flagship

TMCS — Highway Incident Detection

Incident detection on live PTZ cameras: dual-model detection, occlusion-resilient tracking, 7 incident rules and tamper-evident evidence for the ATMS control room.

Pipeline p50
1.5 ms
Rules
7
Tests
298+
  • YOLOv8
  • RT-DETR
  • ByteTrack
  • FastAPI
  • PostgreSQL
Case study Overview PDF
03Production

Vehicle Re-Identification & Forensic Matching

Multi-modal vehicle retrieval combining ResNet + OSNet embeddings, OCR'd plates and vector search. FAISS IVF indexing made retrieval 30× faster, serving 150K+ forensic queries a day at 99.9% uptime.

Search
<100 ms
Queries/day
150K+
Speed-up
30×
Uptime
99.9%
  • PyTorch
  • YOLOv5
  • FAISS IVF
  • OSNet
  • ResNet
  • Redis
  • FastAPI
Code Case study
04Open source

Data Annotation Platform for CV

Local-first dataset platform for model-assisted annotation: YOLOv8/11 auto-labelling, box and polygon instance segmentation, 5-point dataset QA and in-browser fine-tuning with live SSE logs. Auto-annotates 100,000 images in 30 minutes on a Jetson Orin NX 16 GB.

Throughput
100K / 30 min
Labels
Box + polygon
  • Python
  • Flask
  • YOLOv8
  • YOLO11
  • PyTorch
  • Orin NX
Code Case study
05New

CivicMind

Decision-intelligence platform acting as a "smart city brain": multi-agent orchestration (Gemini 2.5) to assess risk, a digital-twin simulator, and explainable AI for transparent policy recommendations.

AI agents
4
Explainable
XAI
  • Gemini 2.5
  • Vertex AI
  • FastAPI
  • Next.js 14
  • Docker
Live Code Case study
06Live

OmniSight

Multi-modal incident intelligence fabric. Correlates high-velocity signals from IoT, cameras and microphones with an agentic workflow engine to validate anomalies and generate safety guardrails in real time.

Inputs
Multi-modal
Engine
Agentic
  • FastAPI
  • Agentic AI
  • RAG
  • Vector DB
  • Docker
Live Code Case study
07New

DocuAI Studio v3.1

Document AI + RAG platform with hybrid retrieval (FAISS + BM25 + RRF), streaming Q&A, a knowledge graph explorer and analytics. Handles PDF, DOCX, XLSX, images and web URLs with sub-2 s answers.

API endpoints
40+
Retrieval
Hybrid
  • LangChain
  • FAISS
  • PaddleOCR
  • FastAPI
  • Ollama
  • Docker
Code Case study
08Live

KeyFlow AI — 10-Agent Typing Studio

Local-first adaptive typing trainer built on a 10-agent deterministic AI architecture with no cloud dependency: sub-millisecond telemetry, O(1) DOM updates, ghost racing and Web Audio synthesis, shipped as an 18 MB standalone executable.

Agents
10
Latency
<1.2 ms
Binary
18 MB
  • Python 3.11
  • pywebview
  • JavaScript
  • SQLite WAL
  • Web Audio
  • PyInstaller
Live Code Case study
09Live · Adopted

Alumni Management Portal

Role-based alumni networking platform, officially adopted by the college and serving 500+ alumni. Four dashboards (Alumni, Mentor, Coordinator, Admin), JWT auth, XSS/CSRF protection, 164+ commits.

DB models
19
REST APIs
50+
Dashboards
4
  • Django 5.2
  • DRF
  • PostgreSQL
  • JWT
  • ES6
Live Code Article
10AIR 1 · SIH 2023

E-Zero — AI E-Waste Management

National-champion system from Smart India Hackathon 2023: IoT smart bins with AI device identification, GPS tracking, weight and fill-level sensing, a real-time analytics dashboard and a citizen mobile app.

Result
AIR 1
Cost
−40%
Efficiency
+60%
  • ESP32
  • Django
  • AWS IoT
  • MQTT
  • React
Live Article
11Live

JalTantra — Smart Irrigation

IoT irrigation with ML-driven water optimization from real-time soil-moisture sensors and weather APIs. University Rank 2, SPPU Startup Olympiad 2025.

Water saved
30%
Olympiad
Rank 2
  • Python
  • Raspberry Pi
  • Scikit-learn
  • AWS IoT
  • Flask
Live Overview Hackathon story

06 Toolkit

End-to-end AI — from training a model to running it on the edge.

Computer vision

  • YOLOv5 / v8
  • RT-DETR
  • OpenCV
  • ByteTrack
  • Kalman filtering
  • Multi-object tracking
  • OSNet
  • ResNet
  • ANPR
  • PaddleOCR PP-OCRv4

GenAI & LLMs

  • LangChain
  • RAG pipelines
  • Hybrid retrieval (BM25 + RRF)
  • FAISS
  • ChromaDB
  • Agentic AI
  • LangGraph
  • DSPy
  • Multi-agent systems
  • Google ADK
  • vLLM
  • Ragas evals
  • Prompt engineering
  • Hugging Face
  • OpenAI API
  • Gemini
  • Ollama
  • Document AI

ML frameworks

  • Python
  • PyTorch
  • TensorFlow
  • Scikit-learn
  • NumPy
  • Pandas
  • NLTK
  • MATLAB

Inference & edge

  • TensorRT FP16/INT8
  • ONNX Runtime
  • CUDA
  • NVIDIA Jetson Orin NX / Nano / Xavier
  • GStreamer (nvv4l2decoder)
  • TensorRT-LLM
  • Raspberry Pi
  • Rockchip RK3588
  • ESP32
  • Arduino

Backend & streaming

  • FastAPI
  • asyncpg
  • Flask
  • Django / DRF
  • REST
  • Microservices
  • RTSP
  • ONVIF
  • MJPEG
  • Server-Sent Events
  • Socket.IO
  • MediaMTX
  • systemd

Data

  • PostgreSQL
  • Redis
  • MongoDB
  • MySQL
  • MSSQL
  • SQLite (WAL)
  • SQL

Cloud & MLOps

  • Docker
  • Kubernetes
  • CI/CD
  • Git
  • Linux
  • AWS
  • Azure
  • GCP
  • Vertex AI
  • Triton Inference Server
  • Prometheus

Frontend

  • JavaScript (ES6+)
  • React
  • Next.js
  • Node.js
  • HTML5 Canvas
  • CSS3
  • Bootstrap

07 Recognition

National and global hackathon wins.

AIR 1

Smart India Hackathon 2023 — National Champion

Team Lead · AI E-Waste Management System

Led the team to national victory with an AI system for automated waste classification — selected as the best solution from thousands of teams across India.

Rank 2

SPPU Startup Olympiad 2025

University level · AI Smart Irrigation (JalTantra)

Predictive ML for water optimization — IoT and AI integration for precision agriculture.

Runner-up

TIAA Global AI Hackathon

Global · AI Investment Intelligence

Risk-based scoring and data-driven financial recommendations for AI portfolio insights.

Certifications recent · verifiable

Foundations

  • Career Essentials in Generative AIMicrosoft + LinkedIn Learning
  • Deep Learning OnrampMathWorks
  • Statistics OnrampMathWorks
  • Python ProgrammingGUVI
  • Geodata Processing using PythonISRO
  • GNSS — Satellite NavigationISRO
  • Advanced ExcelMKCL
  • Industrial ElectronicsYash Infotech

08 Writing

Notes from the build log.

All articles

09 Contact

Let's build something that works in the real world.

Open to AI / computer vision roles, projects and collaborations. I usually reply within 24 hours.

Sends straight to my inbox — I usually reply within 24 hours.