NNC

Newest first

Projects

Things I tried. Some worked. Most taught me something.

  1. 13AgentronicsACTIVEWebMCP infrastructure for AI agents.
    Problem
    Agents can now act on websites, and the infrastructure that makes that safe doesn’t exist yet: who the agent is, what it did, what it remembers, and what it’s allowed to see.
    What I built
    WebMCP infrastructure: auth, observability, memory and context management for AI agents working on websites.
    How it went
    Active. The one I’m working on right now.

    WebMCP · TypeScript · Agent infra · Observability

  2. 12Planetary Alignment SearchDONESearching a million years of orbital mechanics.
    Problem
    Finding the tightest planetary alignment over geological time is a brute-force problem that brute force can’t afford.
    What I built
    A two-stage search over JPL Keplerian elements: a coarse sweep to find candidate windows, then refinement inside each one.
    How it went
    Best alignment in a million years: 8.581°, on 20 September 263,486 CE.

    JPL Keplerian elements · NumPy · Hierarchical search

  3. 11MSME Financial HealthDONEPredicting small-business financial distress.
    Problem
    MSME lending decisions get made on thin, messy financials where the signal is in the ratios, not the raw fields.
    What I built
    A CatBoost/LightGBM blend over 24 engineered features, checked with 10-fold stratified cross-validation.
    How it went
    0.8940 Micro-F1.

    CatBoost · LightGBM · Feature engineering · Python

  4. 10pg_ai_queryTESTINGNatural-language querying for PostgreSQL.
    Problem
    Text-to-SQL inside the database is either a trust problem or a latency problem, and usually both.
    What I built
    A sidecar that keeps the model out of the database process, with Context7 MCP for schema context and a benchmark harness to measure accuracy.
    How it went
    Still being tested. Became the basis of a GSoC 2026 proposal.

    PostgreSQL · Context7 MCP · Python

  5. 09Multimodal ChatbotDONEImages and text in one conversation.
    Problem
    Most assistants go blind the moment a question involves an image, which is exactly when people want to ask one.
    What I built
    A chat system that takes images and text in the same thread and keeps context across both.
    How it went
    Working and deployed.

    Vision-language model · Python

  6. 08Compliance RAGDONESearch over regulatory documents.
    Problem
    Regulatory text is long, cross-referenced and painful to search. The answer is in there somewhere.
    What I built
    Retrieval that splits documents along their clause structure and cites the clause each answer came from.
    How it went
    Working and deployed.

    RAG · Vector search · Python

  7. 07ZeraPortfolioLIVEMulti-agent advisor for Indian stock portfolios.
    Problem
    Retail portfolio advice is generic because it never sees your actual holdings.
    What I built
    One agent per stock, reading live Zerodha holdings through MCP and checking them against macro and geopolitical news. Sends a PDF report to WhatsApp at 15:30 IST every trading day.
    How it went
    Running on real holdings.

    Google ADK · Gemini 2.5 · Zerodha MCP · WhatsApp API

  8. 06AutoResearchClawDONEA pipeline that wrote and submitted its own paper.
    Problem
    Research has a long tail of mechanical work (literature sweeps, experiments, tables, writing) that isn’t the idea itself.
    What I built
    A 23-stage pipeline from question to submitted manuscript. Every stage is saved and can be inspected, so the output can be checked instead of trusted.
    How it went
    Produced ECHO, a study of architectural bias in wage-deflation forecasting, which was submitted.

    Multi-agent · LLM orchestration · Python · LaTeX

  9. 05TriageAIDONEMulti-agent support for medical triage.
    Problem
    Triage decisions get made in seconds by tired people, and ranking a patient one level too low can’t be undone.
    What I built
    Specialist agents on Google ADK that argue over a case, with an XGBoost risk model trained on 20,000 real clinical records. It shows where they disagree instead of hiding it behind one score.
    How it went
    NSCIF 2026 finalist, out of 1,000+ teams.

    Google ADK · XGBoost · Multi-agent · Python

  10. 04UIDAI Aadhaar LifecycleDONELooking at how people enrol and update Aadhaar.
    Problem
    Enrolment and update patterns across India’s ID system say a lot about the population, and nobody reads them together.
    What I built
    An analysis of groups of people from enrolment through updates, across the national dataset.
    How it went
    Finished analysis over the national dataset.

    Python · pandas · Statistics

  11. 03Gemma 2B / QLoRADONEFine-tuning a 2B model on one GPU.
    Problem
    Fine-tuning is assumed to need a cluster. I had one P100 and wanted to see how far it went.
    What I built
    4-bit LoRA adapters on Gemma 2B: data prep, training and evaluation, all inside one card’s memory.
    How it went
    It trained. Most of the work was fighting library versions, not machine learning.

    Gemma 2B · QLoRA · PEFT · bitsandbytes

  12. 02Medi Pro SolutionsSTOPPEDNABH certification help for hospitals.
    Problem
    Small hospitals want NABH accreditation but can’t make sense of the standard or the paperwork it needs.
    What I built
    A small agency: outreach, a service model, and a step-by-step process from first audit to submitted file.
    How it went
    Stopped. Learned that B2B services run on trust built over months, not on a website.

    B2B · Compliance docs · Outreach

    Live ↗

  13. 01FlauraSTOPPEDFlowers straight from the farm to the customer.
    Problem
    Flowers pass through three middlemen and a day on a shelf. By the time they arrive, they’re already dying.
    What I built
    A delivery service with one hop from farm to customer and fixed morning and evening slots, so stock was picked for real orders.
    How it went
    Stopped for lack of money. Learned that timing and capital matter as much as the idea.

    Ops · Landing page · Manual delivery

Also happened

Say hi