Hello.
|
Intro
i’m soumya. i design, i build, and i mess with things just to see how they work. spent a couple years in startup ops & strategy, but recently i've put most of my energy in the deep end: ai tooling, agents, frontier coding tools, and rapid prototyping.
i'm a compulsive tinkerer. obsessed with the entire process of how things are built, curious about any and everything. i dig into everything: cloud, infra, webdev, ios, ml and everything in between.
Experience
Komma Financial Technologies
Nov 2022 – Jan 2024Simplifying investing by developing personal finance tools that make managing money easier.
- Worked directly with early-stage founders to drive day-to-day execution across operations and product, standardizing internal SOPs.
- Owned the translation of business and financial requirements into product flows, collaborating closely with engineers to ship website and internal tool updates.
- Spearheaded a cross-functional team of interns across analytics, marketing, and fintech research, setting workstreams, reviewing outputs, and driving execution against clear deliverables.
BabyKavach
Jan 2024 – Jan 2025Building a suite of digital healthcare services to help parents with vaccinations and medical records.
- Conceptualized and managed the roadmap for components, translating complex health metrics into actionable user insights in React.
- Worked on Go-to-Market strategy for the brand’s digital presence; managed ad budgets and content operations to drive initial user acquisition and engagement.
- Worked the end-to-end development of the brand website and application, optimizing the User Journey (UX) to serve as the primary lead generation funnel.
selected works
A collection of experiments, agents, and side-projects.
my personal agent
Eclipse
Powered by Hermes. Runs 24/7 on a server — I talk to it through WhatsApp, Telegram, or a web portal depending on where I am. It has long-term memory across every conversation, ingests my Gmail, Whoop recovery data, Strava activity, and Swiggy orders — and uses all of it to actually know me. It knows my projects, my portfolio, my schedule, my patterns. Not a chatbot, not a tool — closer to a persistent digital version of me that's always on.
side projects
Upload any company's annual report PDF — Fulcrum extracts ~25 financial fields via Gemini, computes Tier-1A ratios, runs two ML models in parallel (trained on a cohort of wilful defaulters vs. controls), and synthesizes a structured analyst memo, streamed to the UI in real time. Built to replace manual ratio checks in Indian credit underwriting with a grounded, explainable scoring pipeline.
Snap a meal, and Glai runs a two-step OpenAI analysis — first identifying dishes, then estimating weights and nutrition ranges. Data is saved locally in SQLite with optional Supabase sync. Built for a single-user diabetes use case, with HealthKit/CGM integration planned. Shipped to TestFlight via EAS.
SHIT
[Student Helper Integrated Tool] A minimal native app for managing academic timetables and tracking attendance. All data stored locally on-device with no cloud dependencies. Features weekly timetable management, visual attendance tracking, and insights.
A Smart India Hackathon (SIH) project developed for Kochi Metro, focusing on improving operational efficiency by solving ops problems like depot stabling & turnout order, and optimizes safety, service readiness, branding exposure and mileage balancing.
PhotoCortex
PhotoCortex is an AI-powered photo analysis and organization platform built on an ensemble of multimodal AI models to help you understand and explore your photo collection in new ways. It combines face, object, scene, and text detection with computer vision — Google Photos, but fully local.
Developed for SIH 2024 (Bharat Electronics Limited). This system integrates YoloV8 for computer vision and Llama 3.1 for NLP to analyze images and generate contextual responses for component identification and retrieve information about the particular part from BEL's official documentation.
A compact repository implementing AlphaFold data preparation pipelines and Reinforcement Learning algorithms (Actor-Critic CartPole, Q-Learning FrozenLake). Features a dashboard for visualizing training metrics and comparing toy backbone inferences against AlphaFold traces.