HSHarshit Sharma
REC · CH01 ROAD40 KM/H26.82914 N · 75.56997 E

engineer · 0.98Harshit Sharma

I build perception pipelines
that work in the real world.

Co-Founder & CEO of Civilytix, incubated at AIC IIT Delhi and winner of the Honda India Foundation’s Bharat Safety Competition. I lead product, engineering and AI, from dataset to deployment.

1,408road hazards graded
170+ hIndian road footage mined
2 × 2apps live on Google Play and the App Store
3–4×LLM cost cut from model routing
Honda India FoundationBharat Safety winner AIC · IIT DelhiIncubated International Road Federation2026 Start-up Label · Geneva Startup HaryanaSeed Fund DTU SkylabAcceleration Award · Denmark NITI AayogWater Innovation Challenge ISRO PRLResearch presentation IIT KharagpurResearch internship

Shipped products

Two apps live on both stores, built and run by Civilytix with me leading engineering. Every number comes from production code, measured costs or our own road data.

Civilytix Maps

Real-time road-hazard perception for Indian roads, using the dashcams and phones already in vehicles.

Role
Co-Founder & CEO, lead architect
Models
Ultralytics YOLO11s segmentation fine-tuned on Kaggle · Google SigLIP 2 zero-shot
Stack
Python FastAPI · PyTorch · PostGIS · Node.js · React Native · Google Cloud Run
Traction
Paying truck-fleet pilot in Faridabad · Bharat Safety winner · IRF 2026 Start-up Label, Geneva
Perception pipeline
dashcam JT/T 808/1078→1 frame/s while moving→YOLO11-seg · SigLIP 2→blur faces & plates→3-vehicle consensus→live map
  • Wrote an in-house gateway for the dashcam protocols. It streams road frames with GPS, speed and heading, and analyses frames only while the vehicle is moving.
  • Merged several public and in-house datasets into one class map, cut a 45 GB corpus to a 5 GB balanced set, published it as a Kaggle Dataset and fine-tuned YOLO11 on Kaggle’s NVIDIA T4 ×2 GPUs.
  • Mined 170+ hours of road-camera footage with Google’s SigLIP 2 vision-language model on Kaggle. One forward pass scores each frame against 15 hazard prompts and negatives, with resumable checkpoints in Cloud Storage.
  • Model-assisted labelling in CVAT, multi-frame tracking that folds 10–15 detections into one event, and confidence triage that sends uncertain cases to people.
  • PostGIS merges reports within 20 m and publishes a hazard only after 3 independent vehicles agree. Frames that fail anonymisation are rejected before storage.
Civilytix Maps live hazard map with navigation and dashcam feed
Choosing a travel mode: car, motorcycle, bicycle or pedestrian
Privacy consent screen explaining anonymisation
LIVE MAP · TRAVEL MODES · PRIVACY CONSENT
1,408 GRADED HAZARDS FROM OUR OWN ROAD DATA
Stopped vehicle649
Faded lane markings519
Road debris113
Stray animals101
Pedestrians11
Oil patches9
Potholes3
Dashcam frame of a cow on the highway shoulder, detected and graded
Detection · severity 3.00Stray animal at 40 km/h, 26.82914 N 75.56997 E
Harshit in the cab of a truck on the first fleet pilot
First paying pilotIn the cab on day one of our truck-fleet pilot
Hazard map trail coloured by severity along a highway corridor
Live hazard mapEvery dot is a graded detection with photo and GPS

My Magic Pencil

An AI teacher that draws and narrates a lesson on any topic, or on the notes you upload.

Role
Architect and lead engineer, web, Android and iOS
Models
Anthropic Claude and Google Gemini behind one provider interface, routed per request
Stack
TypeScript · AWS Lambda, DynamoDB, Batch (NVIDIA g4dn) · CDK · Capacitor · Amazon Polly
Unit cost
$0.03–$0.11 per lesson, measured
Generation pipeline
topic · PDF · slides · photo→model router→outline→sections, 2 ahead→schema check & repair→voice + whiteboard
  • Model output is forced into a typed schema through tool calling, then validated and repaired with targeted feedback, so a bad response never breaks the player.
  • Lessons stream: the first section plays while later sections are still generating.
  • Routing, prompt caching, pinned model versions and content-hash dedup cut model spend 3–4×.
  • Long jobs checkpoint and resume across Lambda timeouts. Video renders on scale-to-zero NVIDIA GPU instances. A multi-tenant Schools API shares credit pools with role-based access.
  • Payments through Stripe, Razorpay and Dodo webhooks, plus Apple StoreKit 2 and Google Play Billing verified on the server.
Prompt screen asking what to teach, with upload and language options
Generated whiteboard lesson explaining the cosmic balloon analogy
Creator studio listing published lessons and prices
ASK · WATCH THE LESSON · PUBLISH IN CREATOR STUDIO
Agents & LLM systems
EVVOLV.AI · 2025–26

EmaBot, a multi-agent Slack assistant

One agent routes each question to six tool workflows: plain-English SQL over Postgres, Snowflake analytics, document search over Pinecone, web search, charts and Google Drive ingestion. Sensitive tools check access per user.

Related code: Slack data bot →
n8nRAGNL-to-SQLRedis
MANOTR · 2025

Graph-database agent over MCP

A Python LangChain agent that queries and updates an ArangoDB graph through an MCP server, with local embeddings for semantic search. Running models locally with Ollama cut API cost by up to 60%.

View on GitHub →
LangChainMCPOllamaChromaDB
CIVILYTIX · 2025

Tender matching for small businesses

Every morning, 1,000 new government tenders are embedded with Gemini and matched to each bidder’s profile with pgvector on Supabase, for about $0.05 a day. Flask API with JWT auth and row-level security.

PythonpgvectorSupabase

Research

Model design from first principles, measured on held-out data.

Predicted versus actual urban heat island temperature, closely following the diagonal
ROORKEE WATER CONCLAVE 2026 · IIT ROORKEE & NIH

Urban heat islands across Indian districts, from multi-sensor satellite data

I designed a Wide & Deep residual network and averaged three seeds into an ensemble.

0.90R² ensemble
1.39 °CMAE
1.45 Mparameters
Dense 5124 × residual (256→128→128→64)BatchNorm · Swishwide skip3-seed average

Earlier UHI model on GitHub →

Harshit outside IIT Kharagpur's main building
IIT KHARAGPUR · 2024

Edge road-quality detection

YOLOv8 road-defect detection tuned for edge devices, with 25% lower inference overhead and a map dashboard of detections.

View on GitHub →
ISRO PRL · 2023

Deep-learning surrogate for planetary probes

Trained on simulation data to predict particle positions with under 0.01 MSE. Presented at the International Conference on Planets, Exoplanets & Habitability at the Physical Research Laboratory, Ahmedabad.

View on GitHub →

Journey

B.Tech, Computer Science (Data Science), VIT Vellore, 2026 · GPA 8.57

Civilytix Pvt. Ltd. Co-Founder & CEO

Lead product, engineering and AI for Civilytix Maps, Blindspot, My Magic Pencil and the tender-matching platform. Incubated at AIC IIT Delhi, backed by the Honda India Foundation’s Bharat Safety programme and the Startup Haryana Seed Fund.

Evvolv.ai AI & Automation Engineer

Built EmaBot and automation across Slack, Snowflake, Pinecone, WhatsApp, Shopify and WooCommerce.

Manotr AI Engineer

LangChain agent over ArangoDB through MCP, local model inference and vector retrieval.

DTU Skylab, Denmark Next Generation Digital Action

National Winner, International Challenge Winner and Acceleration Award. Presented ML-based bid analysis for water and sanitation infrastructure on stage at the Digital Tech Summit, Copenhagen.

IIT Kharagpur Research Intern, Computer Vision

YOLOv8 road-quality detection for edge deployment.

Recognition
  • Bharat Safety CompetitionWinner · Honda India Foundation × AIC–IIT Delhi · ₹10 lakh grant
  • IRF 2026 Start-up LabelInternational Road Federation · 1 of 5 worldwide, only one from Asia
  • Next Generation Digital Action 2024National Winner & Acceleration Award · DTU Skylab, Denmark
  • Startup Haryana Seed FundBacked
  • Water Innovation Challenge 4.0Winner · NITI Aayog
  • Smart India Hackathon · Nokia Accelerate HerNational Finalist · Finalist
Contact

Let’s build something that ships.

Open to AI engineering, deep learning and computer vision roles. Happy to relocate.