Things I've shipped.

2022 — 2026

ApexOS

Company · Apex Infosys India

Home and building automation at Apex Infosys India.

I worked on the hardware, firmware and cloud services behind our automation products, including custom PCBs, device management and voice control. The systems were used in homes and hotels.

• — Device firmware using ESP-IDF and FreeRTOS, with MQTT communication through AWS IoT Core and over-the-air updates.
• — Voice control using on-device wake-word detection, Whisper, text-to-speech and LLMs.
• — Remote access through TLS tunnels, and Buildroot-based Linux images for device setup and recovery.

2025

InvoiceAI

Research · arXiv 2511.05547

Hybrid table detection + OCR + LLM pipeline with guardrails that stop the model from inventing fields.

Invoices are the worst-case input for LLMs: noisy scans, inconsistent layouts, and fields that must be exact. This work combines classic table-detection models with LLM extraction, then wraps the model in confidence scoring and schema validation — so when the system isn't sure, it says so instead of guessing.

≈95%

field accuracy

−80%

manual review

10k

invoices tested

• — TableNet / CascadeTabNet detection, layout-aware parsing and graph-based entity mapping.
• — Confidence scoring, schema validation and rule guardrails to suppress hallucinations.
• — 30–40% better end-to-end accuracy than OCR-only and rule-based baselines, stable on noisy scans.

2024 — 2025

DDRC Platform

Gov. of India · Ministry of Social Justice

One accessible platform for rehabilitation services for persons with disabilities, from intake to policy dashboards.

India's disability rehabilitation services were fragmented across departments and paper records. The DDRC platform gives each beneficiary a single record that follows them across medical, education and rehabilitation services — built accessibility-first, because its users are the people most failed by typical government software.

WCAG

accessibility-first

1

unified record

IEEE

published

• — Accessibility-first UI for visual, auditory and motor impairments.
• — Brought separate departments onto one platform, with medical, education and rehab records in a single backend and live dashboards for officials and NGOs.
• — ML for beneficiary clustering and demand forecasting to surface regional service gaps.

2024 — 2025

HyperVote

Systems · Security

A permissioned-blockchain e-voting and audit network with private ballots and verifiable tallies.

E-voting fails on trust, not throughput. HyperVote explores whether a permissioned blockchain can give voters cryptographic receipts and auditors verifiable tallies without exposing a single ballot — with the endorsement policies formally verified rather than assumed correct.

500

TPS load-tested

<5s

finality

0

tamper incidents

• — Three-org Hyperledger Fabric network on Raft with private data collections isolating voter PII.
• — AES-256 + RSA-2048 + ZKPs for ballots, homomorphic tallying, verifiable receipts.
• — Zero-trust architecture throughout; led as my senior-year project.
• — Endorsement policies formally checked with TLA+.

2023 — 2024

Landslide Watch

Research · Geospatial ML

CNN change detection on satellite imagery, fused with terrain and rainfall data, to map landslide risk in Raigad.

Raigad district faces deadly landslides every monsoon. This project fuses multi-temporal satellite imagery with elevation and rainfall data to map where the next one is most likely — turning change detection into an early-warning layer for local administration.

DEM

+ IMD rainfall fusion

GEE

multi-temporal

• — Patchified multi-temporal Sentinel-2 imagery from Google Earth Engine as model input.
• — CNN classifier with DEM slope and IMD rainfall fused as auxiliary channels.
• — Hazard heatmaps delivered in ArcGIS for district-level planning.

2020 — 2022

Crophle

Company · Co-founder & CTO

Agri-tech using satellite, market and weather data to give small farmers localised crop and price insight.

Small farmers make high-stakes decisions with almost no data. Crophle combined satellite imagery, mandi prices and weather feeds into localised, speech-first advisories — built for farmers who read prices better than prose.

APMC

+ weather + IoT feeds

• — Satellite, APMC mandi price, weather and IoT feeds fused into localised advisories.
• — Forecasting for demand, pricing and logistics across the supply chain.
• — The hard part was trust: farmers, local vendors and institutional buyers each had to rely on it before the analytics mattered.
• — Multilingual, speech-first UX designed for low-literacy rural users.