AI, backend and full-stack engineer, graduating in Computer Science from the University of Hertfordshire in 2027.
Read live from GitHub, newest first, out of 131 repositories.
You type a rough idea. It interviews you, drafts a prompt tailored to the platform, and scores it against a rubric you can audit.
13agents
orchestrated by a 13-state supervisor state machine
90endpoints
REST, over 29 encrypted tables
7surfaces
web, Chrome and Firefox extensions, Electron, native macOS and an SDK
345+models
routed across Groq, Gemini, Cerebras and OpenRouter for cost and latency
Founder. Built in TypeScript, Next.js, React, Hono, AWS Lambda, PostgreSQL, pgvector, Electron and Swift.
Code on GitHubThen the rest of the 78 I can show you, from 131 repositories, searchable.
Cameras you instruct in plain words, in your own language, with every alert checked by a second model.
An open-source AI steward with a computer of its own. It does the work, asks before it acts, and keeps a record.
A gateway that detects distillation theft of AI models sold through an API, and proves ownership of stolen copies.
A cross-chain treasury agent where the AI makes the decision and the chain enforces the limits.
Clinical photography and research records for surgeons, isolated per user with row-level security.
An autonomous SOC war room: open-source LLMs fold alert storms into incidents and triage them with specialist agents.
Showing 15 of 71 projects not featured above
| Project | Area | Built with | Last push | Links |
|---|---|---|---|---|
| qPilot Open, validated qPCR analysis that shows which of your analysis choices could change the conclusion. Research use only. Reinvent the Wheel hackathon | Research | Python | Code | |
| QuantNature Private Flood Danger Index for an India pilot grid, a Resilience Gap Index ranking 228 countries, evidence-cited action plans and a globe UI. | Research | Python, MapLibre, DuckDB | ||
| Personal Website Private This portfolio with blog, Neon DB, Medium-style editor. | Full Stack | Next.js, Neon, Tiptap | Live | |
| PIO App Private Expo mobile app with a Supabase backend. | Mobile App | TypeScript, Expo, Supabase | ||
| UniPhish Unicode-native byte-level 1D-CNN for phishing URLs across six writing systems, with an honest reproduction of the paper: 0.9954 F1 on 7,199 held-out real URLs. | Research | Python, TensorFlow, TFLite | Code | |
| LAAC Forms Private Forms web app built with React and Vite. | Full Stack | JavaScript, React, Vite | Live | |
| Sarvam Voice Agent Indian-language voice agent on Sarvam's streaming speech-to-text, LLM and text-to-speech; it starts speaking about 0.6 to 1.2 s after you stop. | AI/ML | Python, Sarvam AI, WebSockets | Code | |
| Deepgram Voice Agent Real-time voice agent: Deepgram speech-to-text, FastEmbed RAG, Gemini or Groq for reasoning, and local Piper TTS in English and Hindi. | AI/ML | Python, FastAPI, Deepgram | Code | |
| EasyEdits Gemini app built in Google AI Studio. | AI/ML | TypeScript, Gemini | Code | |
| ComputerVision Orchestra Visual evidence verification system for damage claims across cars, laptops, and packages, comparing submitted images against claim text, history, and evidence requirements. HackerRank Orchestrate | AI/ML | Python, Computer Vision, VLM | Code | |
| GeoShardDB Multi-region distributed database simulator with PostgreSQL shards, FastAPI routing, Redis cache-aside, circuit breakers, Prometheus metrics, and a RAG documentation assistant. | Full Stack | Python, FastAPI, PostgreSQL | Code | |
| AttendAI Anti-proxy attendance platform with admin, teacher, and student dashboards, session management, notifications, reports, and face descriptor support for attendance verification. | Full Stack | TypeScript, React, Express | CodeLive | |
| Decoding Tech Web Private Website for the Decoding Tech community. | Full Stack | TypeScript | Live | |
| NueroAI Demo Gemini app built in Google AI Studio. | AI/ML | JavaScript, Gemini | CodeLive | |
| Decoding Tech Community Private Community platform for Decoding Tech. | Full Stack | TypeScript |
Twelve months and more of industry work: voice AI, telecoms security, a startup platform I led, and clinical ML.
Shipped production generative AI voice agents for customer engagement, lead generation and support.
Infrastructure and Network Security operations at a global telecoms company.
Owned the V1 product from zero to launch; the platform grew to around 10,000 users.
Clinical decision support from electronic health records, built with doctors and medical staff.
Multilingual phishing defence, and voice AI that people who speak Indic languages can afford.
0.9954F1
UniPhish on 7,199 held-out real-world URLs, measured in my 2026 reproduction of the paper.
A byte-level 1D-CNN that reads every URL as UTF-8 bytes, so one model with a 256-token vocabulary covers Latin, Arabic, CJK, Devanagari, Hangul and mixed scripts. In 2026 I reproduced the paper end to end and published where it held and where it did not.
Two of the paper's claims did not reproduce: the parameter count, and that ASCII detectors are blind to non-Latin scripts. The rerun says so openly; the byte-level design earns its place on homoglyph robustness.
0.9991macro-F1
in distribution, 6 writing systems
0.995accuracy
at 20% homoglyph substitution
0.40MB
int8 TFLite, runs on-device in a Chrome extension
0.76ms
per URL on a laptop CPU, one thread
A self-hosted voice agent for Hindi, Tamil, Telugu, Kannada, Tulu and English with automatic language identification, built to cost a fraction of commercial voice APIs.
It also flags emotional distress, for the safety-critical calls a voice agent should never miss.
742ms
end to end, benchmarked on an AMD Instinct MI300X
488ms
time to first sound
85 to 90%
lower running cost (₹0.95 a minute against ₹7 to 15)
6languages
with automatic language identification
37%
lower word error rate for Tulu, a minority language
Over 45 hackathons and CTFs entered. I compete with a CTF team ranked #38 in India.
| Competition | Result | What it took | Record |
|---|---|---|---|
| TraceBash CTF 2026 | 1stglobally | 6,154 points across reverse engineering, crypto, web, forensics and binaries | CTFtime |
| RIFFHACK: Black Market Break-In | 2ndglobally | 5,955 points, CTFtime 2026 | CTFtime |
| HackMIT | FinalistMIT, USA | 33rd worldwide in the admissions puzzle; the top 50 earn entry | |
| Snipers CTF | 1st | Offensive security and problem solving | |
| CompFest CTF | 1st | Practical security analysis and exploitation | |
| BushBash CTF | 2ndAustralian National University | Podium in a cybersecurity competition | |
| IIIT Hyderabad Hackathon | Finalist | Voice AI agent for Apollo Hospitals: patient queries and appointments | |
| IIT Roorkee Frontend Hackathon | Finalist | Sponsor-focused event website | |
| MIT Manipal Hackathon | 4th | iWitness, anonymous crime reporting | |
| NASA Space Apps Challenge | Participant | Data visualisation and simulation on NASA datasets | |
| LA Hacks | SelectedUCLA | Selected to compete | |
| HackGT | SelectedGeorgia Tech | Selected to compete | |
| Big Red Hacks | SelectedCornell University | Selected to compete | |
| UCF Hackathon | SelectedUniversity of Central Florida | Selected to compete | |
| Junction | SelectedFinland | Selected to compete | |
| SinceAI | SelectedFinland | Selected to compete |
In 2026 I moved to the University of Hertfordshire for the final year of my Computer Science degree.
I build full-stack and backend systems end to end, then measure them: a platform grown to around 10,000 users, a 742 ms real-time voice pipeline, a phishing model that runs in 0.76 ms on a laptop CPU.
My first product, a university companion app, was acquired by Colcord, where I went on to lead the V1 platform. Since then I have built voice agents at VoiceDots, worked in Ericsson's Infrastructure and Network Security team, and built clinical ML with doctors at Kasturba Hospital.
My research is on UniPhish, multilingual phishing detection, and affordable multilingual voice agents for Indic languages. Outside coursework I play CTFs and build in the open.
BSc (Hons) Computer Science
B.Sc. (Applied Sciences) Computer Science and Engineering
High school, 93%
Coursera
Notes on software, AI and building products.
I had a very long vacation after the completion of my course at ICAS Manipal, where this vacation was a complete roller coaster moment there were days where I s...
While building apps in this era where its been a norm just to vibe code the project using Claude or Codex, what I noticed is that not many people thinking about...