Hardik Gulati
Third-year B.Tech student in AI & Machine Learning at Symbiosis Institute of Technology, Pune. I build data-driven systems — retrieval pipelines that cite their sources, accessibility tooling that files its own bug reports, document AI, and fraud detection. Focused on turning messy data into decisions that matter.
Projects
-
Scholar
Research assistant that answers only from papers you upload — every claim cites the exact source page. Token-accurate chunking, mpnet embeddings into a per-user FAISS index, and generation constrained to retrieved passages. Retrieval changes ship on measurement, not vibes: pseudo-relevance feedback was benchmarked and rejected, HyDE kept behind an opt-in flag.
FastAPIFAISSsentence-transformersPyMuPDFOllamaReactTypeScriptJul 2026 —
Present -
A11y GitHub Issue Logger
Turns WCAG findings into reviewable GitHub issues in four steps: scan a URL with axe-core, search the repo for duplicates, draft a structured issue, log it. Playwright highlights and screenshots the offending element as real evidence, and authenticated pages can be audited via a one-shot session that is never persisted. Human review is mandatory — the tool drafts, you decide.
FastAPIPlaywrightaxe-coreReactDockerGitHub APIJun 2026 —
Jul 2026 -
InvoiceIQ
Financial document processing for Indian MSMEs, built under the Ministry of Micro, Small and Medium Enterprises. Upload a photo or PDF of any invoice; the system extracts vendor, amount, GST, date, and invoice number, categorises the expense, validates GST details, and feeds a live spending dashboard. Built for real-world documents — handwritten bills, regional-language headers, non-standard layouts.
PythonOCRscikit-learnDockerREST APIPrivate repository — available on requestJun 2026 -
RAD-SCRIBE
Evidence-grounded chest X-ray reporting pipeline that audits its own clinical claims before finalising a report. Vision Transformer + DistilGPT-2 generation layered with FAISS-based RAG, counterfactual evidence retrieval, claim-level verification, and sentence-wise confidence scoring to filter low-confidence hallucinations. Evaluated across 7,000+ real image-report records using BLEU, ROUGE-L, and BERTScore. Team project, presented to the SIT director.
PythonPyTorchViTDistilGPT-2FAISSRAGBERTScoreJan 2026 —
Apr 2026 -
FraudGuard
Banking transaction analyser and fraud detection engine. Spring Boot REST service over MySQL with a Caffeine-backed cache and a JavaFX desktop client, scoring transactions against Logistic Regression and Random Forest pipelines trained on the PaySim dataset. Evaluated using Precision, Recall, and F1-score against severe class imbalance.
JavaSpring BootJavaFXMySQLscikit-learnFeb 2026 —
Apr 2026 -
NexaKind
Smart donation rerouting platform for orphanages — redistributes surplus resources to the shelters that need them most. Role-based authentication, real-time pledge tracking, SQL-driven analytics, and MySQL schema design behind a unified donor/orphanage interface. Built with two collaborators under faculty mentorship.
FlaskSQLAlchemyMySQLPythonTailwindAug 2025 —
Nov 2025
Skills
Experience & Credentials
Volunteer
B.Tech in AI & ML
Class XII — CBSE
Class X — CBSE
Artificial Intelligence With LaunchED
Supervised Machine Learning
Hashgraph Developer Course
Machine Learning Specialization
1st place — Capture the Flag
Level 4 badge — open source contributions
Open to data & AI
roles & internships.
Looking for opportunities to work on real problems with real data. Let's build something meaningful.