Sai Pranav
AVATAR.JPG
I am R Sai Pranav. CSE student at REVA University with hands-on experience in software development, system design, and AI/ML engineering.
> SDE Intern at Cepheid (Danaher) — Siemens Opcenter MES.
> Previously: Full Stack, Cybersecurity & Explainable AI internships.
> 50+ merged patches to PostgreSQL tooling, GRASS GIS, GNU Radio, JabRef & Kubernetes.
> Filed a personal patent; national hackathon winner.
> Competed in 20+ hackathons including Google DeepMind, Cardano Asia, Razorpay & Snowflake.
/// SYSTEM_OPTIMIZED
@ Cepheid India Pvt Ltd (Danaher) — Bangalore
@ PostgreSQL ecosystem (pgmoneta, pgagroal, pgvictoria), OSGeo GRASS GIS, GNU Radio, JabRef, Kubernetes
@ MyStartupWave — Bangalore (Hybrid)
@ iSPIRT Foundation (Sakram) — Remote
@ CSIR–4Pi (Fourth Paradigm Institute) — Remote
/// LIVE_FEED: Pranav-error
Segments 100–2,000 page mortgage packages into 27 document types at 91% accuracy. Cost-aware cascade (pdfplumber → heuristics → GPT-4o-mini VLM) so most pages never reach an LLM; Probabilistic Table Threading (Naive Bayes over 5 signals) auto-stitches ~85% of cross-page tables. Validated on 2,295 pages.
Is this document real? Verifies dispute evidence and GST invoices with cryptographic and arithmetic ground truth (signed QRs, GSTN e-invoice payloads, GSTIN checksums) instead of pixel forensics, which is blind to AI inpainting. Tested on 9 independent corpora: 15 real GSTINs accepted, 0 false positives.
AML copilot that takes an analyst from a fraud signal to a cited, machine-checked SAR entirely inside Snowflake. 5 deterministic rules mapped to policy clauses — the LLM never decides what is fraud. Precision 0.71–1.00 on a held-out dataset; append-only decisions, PII masked even inside AI answers, 36/36 governance checks.
Space-enabled waste intelligence for Bengaluru: Sentinel-2 dump detection (Random Forest, 2/2 ground-truth sites matched), XGBoost risk prediction over a 552-cell grid, and route optimisation cutting ~40% of daily distance. 9 ML models behind a 34-endpoint FastAPI + Next.js platform.
Segments 100–2,000 page mortgage packages into 27 document types at 91% accuracy. Cost-aware cascade (pdfplumber → heuristics → GPT-4o-mini VLM) so most pages never reach an LLM; Probabilistic Table Threading (Naive Bayes over 5 signals) auto-stitches ~85% of cross-page tables. Validated on 2,295 pages.
Is this document real? Verifies dispute evidence and GST invoices with cryptographic and arithmetic ground truth (signed QRs, GSTN e-invoice payloads, GSTIN checksums) instead of pixel forensics, which is blind to AI inpainting. Tested on 9 independent corpora: 15 real GSTINs accepted, 0 false positives.
AML copilot that takes an analyst from a fraud signal to a cited, machine-checked SAR entirely inside Snowflake. 5 deterministic rules mapped to policy clauses — the LLM never decides what is fraud. Precision 0.71–1.00 on a held-out dataset; append-only decisions, PII masked even inside AI answers, 36/36 governance checks.
A personal control plane for Claude Code setups spanning multiple accounts and machines: memory sync over your own git remote, a skill registry, token-usage reporting from session transcripts, and a terminal dashboard. Purely local — never calls the API.
Space-enabled waste intelligence for Bengaluru: Sentinel-2 dump detection (Random Forest, 2/2 ground-truth sites matched), XGBoost risk prediction over a 552-cell grid, and route optimisation cutting ~40% of daily distance. 9 ML models behind a 34-endpoint FastAPI + Next.js platform.
Healthcare diagnostics tool that verifies patient data cryptographically without exposing raw records. Won against 200+ teams nationally.
Financial affordability agent that decides pay-in-full, part-pay, installments, wait, or don't proceed. Not a classifier but a cash-flow simulator: projects a 90-day day-by-day balance against recurring commitments and a minimum-balance safety floor, with every calibration sweep documented.
Routes WhatsApp messages to notify / digest / mute per user across text, image posters and voice notes. Deterministic belief-fusion tier settles 23.6% with no model call; 100% action accuracy on gold labels; cold-run cost cut from $2.80 to $0.76.
RAG system for insurance policy documents: upload PDFs and ask natural-language questions to get clause-level answers. sentence-transformers (all-MiniLM-L6-v2) embeddings over a FAISS index behind a Flask API.
A hardware-enforced authorization layer for AI agents — stops an agent from taking actions the user never authorized, enforced in code rather than trusted to the model.
Neural anti-piracy engine — AI-powered cognitive fingerprinting to identify and protect digital content.
Reproducible pipelines for strong-lensing dark-matter classification (ResNet-18, 0.9996 accuracy), cardiac CT radiomics and symbolic regression — each shipping the verification step that would catch it being wrong, e.g. nearest-neighbour leakage controls.
Solutions to ML4SCI evaluation tasks: gravitational-lensing substructure classification, RenAIssance historical-document OCR and SYMBA symbolic computation.
Interpretable prediction of mitotic activity in breast-cancer histopathology using Explainable Boosting Machines and Neural Additive Models — glass-box models instead of post-hoc explanations.
Deep-learning mitosis detector for histopathology images with RISE (Randomized Input Sampling for Explanation) saliency maps showing which regions drive each prediction.
Breast-cancer segmentation models (U-Net, Attention U-Net, ResNet-U-Net) trained from scratch in PyTorch, alongside MedSAM fine-tuning experiments, with XAI heatmaps for clinicians.
Pilot study on whether routine clinical variables from NHANES 2017–2018 can train classifiers to tell alcohol-associated (AFLD) from metabolic-associated (MASLD) fatty liver disease.
Digital companion for field drug testing: kit-agnostic reading of colorimetric tests on existing phones, producing tamper-evident records that plug into the government's evidence systems.
Geospatial FastAPI service over Google Open Buildings around Cubbon Park, Bengaluru — PostGIS spatial queries, ACID writes and an S2-cell sharding demo.
Measures, then fixes, the accuracy of the public geospatial layers (BBMP/BWSSB/BESCOM KML, OpenStreetMap, river and lake datasets) behind a Karnataka cadastral viewer.
Monitors public GitHub for accidentally pushed secrets via Code Search and the push-event stream. Stores only redacted fingerprints, never uses a key, never auto-contacts anyone — with tests asserting no raw secret ever reaches the DB or logs.
Real-time vehicle telemetry from native C++ to Java over a lock-free SPSC ring buffer with a single JVM-attached dispatcher thread. Producer push() latency (p99 ~400 ns) is unaffected by JVM activity. Zero dependencies.
Distributed system that coordinates computation across independent nodes without centralizing data. Structured logging and benchmarking to track performance and handle node failures gracefully.
Decentralized ledger system with graph-algorithm-based verification to ensure data integrity at scale — geospatial filtering, end-node credibility scoring and cryptographic proofs.
Botanical traceability for Ayurvedic herbs with geo-tagging from collection to final label. Hyperledger Fabric (2 orgs, Raft) with Node.js chaincode, FastAPI + MongoDB, React 19 web and offline-first React Native app.
AI-powered cross-exchange arbitrage trading system on BNB Chain.
Full LMS for a 30-day developer challenge: Next.js web, FastAPI backend, Flutter mobile app, JWT + httpOnly refresh auth, Razorpay payments, Cloudinary storage and sandboxed code execution.
Centralized placement platform for colleges — student profiles, company and job postings, eligibility rules, applications and outcomes in one place instead of spreadsheets.
Two-sided marketplace for transferring running businesses in India — structured listings, AI-assisted valuation, document verification and real-time buyer matching.
Tracks student attendance, flags shortage risk in real time and predicts outcomes before exams. Flask + React with JWT auth.
Website for the .Dev Club — Next.js 14, TypeScript and Tailwind, with MongoDB and Notion as content sources.
Custom cloud-scheduling and resource-allocation experiments on the CloudSim 3.0 simulation framework.
Menu bar app for developers whose Macs keep running out of RAM and disk — live usage readout, quit memory hogs, and clean gigabytes of regenerable dev caches. Tauri v2, idles under 10 MB.
Automated API test suite for a login → get-product flow, built for a CRED interview take-home.
macOS-inspired interactive portfolio built with zero frameworks — windows, dock and apps in vanilla HTML/CSS/JS.
LLM-generated summaries of customer reviews on an e-commerce product page, so shoppers get the overall sentiment without reading every review.
AI chatbot on Next.js App Router and the Vercel AI SDK, with authentication, persistent chat history and shareable conversations.
Student mental-health platform: an AI wellness chat, screening questionnaires, counselling booking, a peer community and a resources hub, with an API test harness for the backend.
End-to-end courier platform: customer and driver Flutter apps, a React admin panel (orders, drivers, pricing, reports), a booking website with live tracking map and a FastAPI + SQLAlchemy backend on Firebase.
Runs a TensorFlow segmentation model over an image dataset in parallel with PySpark pandas UDFs, tuned for CPU-only execution.
Modular sensor-driver layer (heart rate, UV, hydration, posture) streaming readings live to a web dashboard over WebSockets.
Next.js App Router + Tailwind starter, deployed to Vercel — the base for later Next.js projects.
Fine-tuned the open-weight Qwen2.5-3B with LoRA / PEFT and Hugging Face Transformers on a custom instruction dataset built from the MedSAM segmentation work.
/// UNLOCKED_RECORDS
Won 2nd place and a ₹75,000 prize at the InfrX 2026 Hackathon with an AI mortgage document pipeline — 91% accuracy across 2,295 test pages. Team Noobda, REVA University.
Patches merged into pgmoneta, pgagroal, pgvictoria, GRASS GIS, GNU Radio, JabRef and kubernetes/website — mostly C memory-safety and correctness fixes, with 20+ more in review.
Released claude-orchestrator as an installable CLI (pip install claude-orc) with CI-tested releases — memory sync, skill registry and usage reporting for multi-account Claude Code setups.
Won Cardano Hackathon Asia beating 200+ teams with Aura AI Diagnostics — a cryptographic healthcare verification tool. Took home a $2,000 (~₹1.8 lakh) prize.
Filed patent for Decentralized Spatial-Temporal Data Veracity System — a distributed ledger with graph algorithm-based tamper-proof verification.
Reached the finals of Google DeepMind Hackathon with Architrust — a distributed, privacy-preserving computation system.
Top 5 at AWI SpaceTech Hackathon (Mar 2026) with OrbitClean 2.0 — Sentinel-2 dump detection, XGBoost risk prediction and route optimisation for Bengaluru.
Member of Google Developer Groups Bangalore (Feb 2026–) and AWS Cloud Club REVA (Mar 2026–) — tech talks, community builds and cloud workshops.
Technical Lead at Algorand Club, REVA University. Ran workshops and smart-contract projects. Organised 5+ hackathons and competed in 20+ including Google DeepMind, Cardano Asia, Adobe India, HackRx, Razorpay and Snowflake.
/// BATTLES_FOUGHT: 20+
Risk Copilot — AML copilot on Snowflake Cortex
SIH26231 — Field drug testing companion (MHA / NCB)
Buy or Wait? — affordability agent
Evidence Verifier — Track 02 AI Risk Manager
WhatsApp Message Notification Router
Mortgage doc pipeline — ₹75,000 prize
Aura AI Diagnostics — $2,000 prize
Architrust
OrbitClean 2.0
Scorpio AI — Anti-Piracy Engine
Botanical traceability on blockchain
—
AI Insurance Document Analyzer
Code Tribe
/// FREELANCE_MODE: ON
I take on freelance builds — full-stack apps, AI integrations, and websites that don't look like everyone else's.
AI contract review and negotiation platform: clause-by-clause risk analysis, ambiguity detection, negotiation strategy, redlined DOCX export with tracked changes and Word comments, and follow-up Q&A. Built the entire backend.
Business website with service catalogue and lead-capture, built and delivered end-to-end.
AI-powered question paper generator that builds exam papers from syllabus input, with queued generation jobs, live progress over sockets and PDF export.
"Sai Pranav proved that no matter the pressure or how little time there is, he completes projects outstandingly."
"Fast, reliable, and actually has good taste in design. A rare combination."
"Cleanest code I've seen in years. He knows how to handle complex state management."
"Creative design ideas and a unique UI experience. The attention to detail is flawless."
"Highly intuitive UX. Delivered exactly what we needed before we even knew we needed it."
"Sai Pranav proved that no matter the pressure or how little time there is, he completes projects outstandingly."
"Fast, reliable, and actually has good taste in design. A rare combination."
"Cleanest code I've seen in years. He knows how to handle complex state management."
"Creative design ideas and a unique UI experience. The attention to detail is flawless."
"Highly intuitive UX. Delivered exactly what we needed before we even knew we needed it."
I am currently available for freelance work and open to full-time opportunities.