$ whoami
Kanishka Chahar
M.Sc. Mathematics (Computing) @ IIT Delhi. Turning abstract numbers into actionable business decisions.
// about me
The Intersection of Math & Data
I believe data is the key to real progress. While I began with a love for Mathematics, it was Statistics that showed me how to turn abstract numbers into practical, real-world solutions.
My purpose became clear during a research project on System Reliability Analysis — building models to calculate component risks and estimate maintenance costs. I realized I wanted to directly influence important business decisions through data.
This led me to pursue my M.Sc. in Mathematics (Computing) at IIT Delhi, where I'm mastering advanced statistical methods and machine learning.
class Kanishka:
education = "M.Sc. Math (IIT Delhi)"
undergrad = "B.Sc. Hons Math (MUJ)"
gpa = {"iitd": 7.4, "muj": 9.61}
strengths = [
"Deep Analytical Thinking",
"Statistical Modeling",
"Visual Communication",
"Leadership & Discipline"
]
seeking = "Data-driven roles"
status = "Open to opportunities"// skills
Technical Arsenal
Languages & Tools
Libraries & Frameworks
APIs & Cloud
Data Skills
Mathematics
// projects
Projects
Built during my internship at India's largest logistics company — 5 interconnected data pipelines that produce a production-grade toll plaza master database (6,500+ plazas, validated coordinates, pricing for 8 vehicle categories). Designed to power freight index price prediction, toll-aware route optimization, and a public-facing website where users can look up toll prices for any journey or plan routes with toll cost estimates.
Document Intelligence | Production System
Toll Receipt Extraction & Validation Pipeline
Professional Project • Jun 2026
Production pipeline processing 33,000+ fleet driver toll reimbursement documents per month — extracting structured transaction data from FASTag bank statements (16+ format variants) and physical toll slip photos using PDF table extraction and Vision AI.
- ▹Built 16+ document format parsers handling 8 different bank statement layouts with multi-page continuation logic
- ▹Integrated Gemini 3.1 Flash Lite for scanned/photo documents — 13,406 files processed with token-bucket rate limiting and exponential backoff
- ▹Vertex AI batch prediction support for 50% cost reduction ($9 vs $17 for 13K documents) — async processing with job status polling
- ▹Crash-safe JSONL streaming architecture: pipeline resumes from exact point of interruption, no duplicate processing
- ▹4-tier toll plaza matching system: plaza code → exact name → compact name → fuzzy match (85% threshold) — mapped 99.98% of 638K transactions
- ▹Direction detection algorithm pairs single/return trips within 24-hour windows using sliding window logic
- ▹Vehicle cross-validation compares extracted VRNs against filename metadata for confidence scoring
- ▹Processing manifest with MD5-based change detection — only processes new/modified files on re-runs
- ▹Total output: 638,275 transactions extracted, 1,437 unique plazas identified, 7.3% flagged for human review
ML + LLM | Hackday Build (8 hours) | Best Approach Award
Vehicle Feasibility Engine
Internal Hackday • Jun 2026
Given any origin-destination pair in India, determines which of 11 truck types can realistically serve that route — using a 3-tier statistical model derived from 323K shipment records and 1.6M historical transactions, enhanced with Gemini reasoning.
- ▹3-tier evaluation strategy: exact OD match (22% of routes) → similar route matching (±30% distance) → statistical p5-p95 distance profiling
- ▹Strict vehicle mapping: 60+ name variants → 11 standard types with zero-ambiguity rule (rejected 20% of data rather than guess)
- ▹Data-inferred city restrictions: detected 78 cities with truck bans from absence patterns in placement data (not LLM hallucinations)
- ▹Statistical distance bands derived from real data: each truck type gets a feasible range (p5–p95) and optimal range (p25–p75)
- ▹Gemini AI generates route-specific explanations without changing feasibility decisions — $0.57 total for 1,451 routes
- ▹Results: 97.3% recall, average 3.0 recommendations per route (tight, not all 11), 2.7s runtime for full test set
- ▹Built complete with Streamlit dashboard, validation pipeline, and knowledge base preprocessing in a single 6.5-hour session
Data Quality Engineering | Multi-Source Validation | EDA
Toll Benchmarking & Enrichment
Professional Project • Jun–Jul 2026
Cross-referencing a 6,500-plaza toll database against TollGuru, Google Places API, and 24,500 live FASTag charges to validate coordinates, detect missing plazas, correct pricing errors, and produce per-source accuracy metrics — with deep EDA diagnostics.
- ▹Pipeline 01: Route-level comparison across 59 OD pairs — greedy 1:1 spatial matching (300m radius), close-loop toll expansion, corridor width sweep (100/200m)
- ▹Pipeline 02: Google Places API validation of ~2,400 plazas by coordinate source (NHAI, OSM, IDFC, TollGuru) — classifies as VALIDATED, COORD_DRIFT, NAME_MISMATCH, or WRONG_LOCATION
- ▹Pipeline 03: FASTag EToll pricing analysis with smart price matching (single/return/CVD detection) and confidence scoring (HIGH/MEDIUM/LOW) based on transaction volume
- ▹Full EDA pipeline (9 diagnostic sections): price error by source, direction-of-error analysis, close-loop distance issues, state-wise accuracy, per-vehicle-class validation, inconsistent rate detection
- ▹Trust hierarchy framework: NHAI (gazette) > TollGuru (live API) > inferred/internal — used to diagnose root cause of pricing discrepancies
- ▹API response caching with deterministic hash keys — $103 one-time Google Places cost, all re-runs free with zero duplicate calls
- ▹Waypoint chunking algorithm: handles DLV Route API's 20-waypoint limit by splitting routes into overlapping windows and stitching geometries
- ▹Plaza detection recall: 90% at 200m corridor | EToll pricing accuracy: 73.6% exact match, 85.3% within ±5%
- ▹Interactive Leaflet maps for every analysis — route selector, colour-coded classifications, per-plaza popups with all metrics
Computer Vision | Prompt Engineering
Toll Price Extraction from Rate Board Photos
Professional Project • Jul 2026
Automated extraction of structured toll pricing data from 498 photographs of Indian highway toll rate boards — multilingual (6+ languages), varied quality, diverse layouts — using Gemini Vision API with carefully engineered prompts.
- ▹99.8% success rate (497/498 images) — handles Hindi, Telugu, Marathi, Gujarati, Kannada, Tamil boards
- ▹10KB engineered prompt with explicit multilingual mapping tables, layout pattern descriptions, and confidence-preserving null rules
- ▹Normalized output: 9 vehicle categories × 4 journey types per image, with strict null vs '0' semantics (unknown vs explicitly free)
- ▹Resume-capable pipeline: --resume flag skips completed images, safe to interrupt and restart
- ▹Total cost: $0.75 for 498 images using gemini-3.1-flash-lite ($0.0015/image) — 49 minutes processing
- ▹Exponential backoff retry strategy with early-abort (3 consecutive failures → stop, prevents wasting API quota)
- ▹Schema validation + normalization ensures every output has identical structure regardless of what the model returns
Data Pipeline | Multi-Source Integration | EDA
Toll Pricing & Expressway Data System
Professional Project • May 2026
Comprehensive toll pricing database for Indian expressways and highways — combines FASTag transaction data (224K records), NHAI official records, TollGuru API, OSM polylines, and Google Places into a unified master dataset with per-km pricing, EDA, and interactive review tools.
- ▹Unified 4,458-plaza master dataset merging 6 independent sources with source-priority resolution (NHAI > IDFC > TollGuru > OSM)
- ▹OSM expressway polyline extraction: graph-based single-carriageway chaining for 27 expressways — avoids ramps and parallel roads automatically
- ▹Rate-per-km computation for both flat tolls (NHAI tollable length) and closed-loop tolls (API road distance between entry-exit pairs)
- ▹EDA on 224K FASTag transactions: price distributions, inconsistent rate detection, expressway entry-exit pair identification, vehicle class patterns
- ▹Identified 196 toll plazas in FASTag data that are NOT in the official NHAI database — discovered via transaction pattern analysis
- ▹Interactive Flask + Leaflet review tool: map-based pricing type classification (flat/close_loop/remove) with real-time backend persistence and CSV export
- ▹Google Places discovery: searches 'toll plaza' every 15km along expressway polylines, deduplicates within 300m of known plazas
- ▹Gemini AI validation for ambiguous name matches between sources — handles spelling variations and regional language differences
- ▹API chunking: handles 20-waypoint limit by splitting routes into overlapping windows with geometry stitching for 27 expressways
// achievements
Recognition & Awards
// publications
Research & Publications
Reliability and Signature Evaluation of Juice Clarification Systems of Sugar Industry
Ashish Kumar, Kanishka, Monika Saini
International Conference on Information and Communication Technology for Intelligent Systems (ICTIS 2025)
Lecture Notes in Networks and Systems (Vol. 1525), Springer Nature
Modeled a complex 8-subsystem industrial Juice Clarification unit as a quantitative reliability problem — transforming real-world manufacturing risk into actionable maintenance insights. Applied reliability block diagram techniques and probabilistic methods (Owen's and Boland's approaches) to evaluate system signature, tail signature, minimal signature, and the Barlow–Proschan Indicator. Computed Mean Time to Failure (MTTF) and expected cost rate to enable data-driven decisions on predictive maintenance scheduling, reducing unplanned downtime and optimizing operational expenditure in industrial settings.
// experience
Where I've Worked
Intern – Data Services (Maps & Toll Systems)
@ Delhivery Limited, Gurgaon
- ▹Built India's most comprehensive toll plaza master database — 6,500+ plazas with validated lat/longs, per-category pricing (8 vehicle types × 4 journey types), expressway metadata, and source provenance tracking across 35+ columns
- ▹Designed to power freight index price prediction, toll-aware route optimization, and a public-facing website where users can look up toll prices or plan journeys with cost estimates
- ▹Engineered 5 interconnected pipelines feeding the master DB: FASTag document extraction (638K transactions) → pricing validation → Google Places coordinate verification → coverage gap detection → confidence scoring
- ▹Built a production document intelligence system (33K docs/month, 16 formats, Vision AI) and a computer vision pipeline for toll rate boards (6+ languages, 99.8% accuracy) — both serve as pricing data sources
- ▹Recognized as 'Best Approach' in internal hackday: vehicle feasibility engine with 97.3% recall using 323K shipment records + Gemini reasoning — built end-to-end in 8 hours
@ Chegg India
- ▹Solved high-volume undergraduate and advanced-level mathematics problems with high accuracy
- ▹Delivered structured, well-documented solutions under strict quality benchmarks
- ▹Consistently maintained strong quality ratings across assignments
Graphic Designer
@ AIC-MUJ (Atal Incubation Center)
- ▹Designed digital creatives using Adobe Photoshop and Premiere Pro for institutional branding
- ▹Collaborated on communication and marketing material for incubation initiatives
@ Manipal University Jaipur
- ▹Led the university mathematics club, overseeing academic initiatives and problem-solving sessions
- ▹Coordinated events, peer-learning discussions, and inter-student collaboration
- ▹Strengthened student engagement and mathematical culture on campus
// certifications
Courses & Certifications
Diploma in Data Science
IIT Madras
Completed first semester with GPA 9/10. Exited to pursue M.Sc. at IIT Delhi.
Introduction to Programming with MATLAB
Vanderbilt University (Coursera)
Programming fundamentals using MATLAB
Introduction to Graph Theory (With Honors)
UC San Diego (Coursera)
Completed with Honors distinction
Data Science / ML Mentorship Program (DSMP 2.0)
CampusX
Comprehensive data science and machine learning mentorship
// contact
Let's Connect
I'm actively seeking data-driven roles where I can contribute from day one. Open to full-time Data Scientist and Data Analyst positions.