Dhanushragav S
Résumé

வணக்கம்! Hello from Pune.

Dhanushragav
Shanmugam

AI engineer, forward deployed at Haqdarshak in Pune

I build voice-first AI agents for India's welfare sector. Before I write the code, I sit with the field workers who'll use it. After I ship it, I keep an eval suite running so I know it still works.

See what I've built Get in touch
    A reconstructed walk-through of each system's real steps. It isn't live data.

    Work at Haqdarshak

    Haqdarshak helps people in India find and apply for government welfare schemes, mostly through trained field workers. I build the AI systems they and their customers use. Each story below opens up into the problem, what I built and how I measured it.

    HD Saathi

    Voice assistant for field workers / May 2026 to now / sole developer

    I pitched this as an in-house replacement for a vendor's chatbot. It teaches workers the schemes, quizzes them, answers spoken questions while they're standing with a family, and picks up training where they left off.

    97% of the 158 eval cases run so far pass
    Read how it was built

    The problem

    Field workers get stuck in the middle of an application, with a family waiting, and the chatbot they had wasn't built around how they actually work. I went out and met them to see where they got stuck, and turned what I heard into the API design and a data model split across Firestore and MySQL.

    What I built

    Three agents talking over A2A, each with search tools: one teaches schemes, one runs adaptive assessments, and one answers voice questions on the spot. They use RAG over 5,000 active scheme records plus a database of app data and FAQs. I wrote the frontend in plain HTML, CSS and JavaScript, and the FastAPI services in Python and SQL against the company database, deployed on GCP.

    How I know it works

    I replaced my manual Python checks with a 186-case eval suite on Google ADK. An LLM judge scores each answer, and code checks confirm the reply language, the session state, and that no Aadhaar or PAN number ever leaks. Then 90 field workers and managers tested it, and their feedback went straight back into the product. It's now being integrated into HD App, Haqdarshak's main worker app.

    • Google ADK
    • A2A
    • RAG
    • FastAPI
    • Firestore
    • MySQL
    • GCP
    • Voice

    Yojana Saathi

    Citizen assistant for finding and booking schemes / Jan to Apr and Jul to Sep 2026

    A citizen asks what they qualify for, and the assistant finds the scheme, checks eligibility and books the appointment. I built the first 80% of it, handed it off, then took it back and made it about twice as fast.

    2× faster replies after I took it back
    Read how it was built

    The problem

    Scheme information is spread across policy documents and several databases, and a booking is only useful if the person is actually eligible. The answer had to be correct and come back fast enough that people don't give up.

    What I built

    A master agent that routes across four sub-agents, using an LLM I fine-tuned on Vertex AI. It runs parallel search over four databases and RAG over 10,000+ policy records, with a React frontend and a FastAPI backend on Cloud SQL, shipped with Docker on GCP. I linked the booking API to RunDQ so eligibility is checked before every booking, used AWS S3 to show uploaded proofs and booking details, and added voice input.

    What changed

    When the consultant I'd handed it to left, I took the product back and fixed production-test bugs with another developer. Context compaction, a tuned thinking budget, caching and a cap on LLM calls brought response time down a lot. It passes a similar ADK eval suite of rubric and code checks at 96%.

    • Vertex AI fine-tuning
    • Multi-agent
    • RAG
    • React
    • FastAPI
    • Cloud SQL
    • Docker
    • AWS S3

    Proof Agent

    Automated checking of field-uploaded proofs / Oct 2026 to now

    My manager built a service that reads about 1,000 uploaded proofs a day with Gemini 2.5 Pro, matches them against MySQL records and gives each one a verdict someone can audit. I'm extending it with a new verification persona.

    Read what I'm adding

    My part

    The new persona gets its own extraction schemas, fuzzy matching for names and fields that never come out quite the same twice, and a set of mandatory rule checks.

    The design rule

    If any mandatory check fails, the proof goes to a human reviewer. The model never gets the final say on a failure, since a wrong rejection can cost a family their benefit.

    • Gemini 2.5 Pro
    • Structured extraction
    • Fuzzy matching
    • MySQL
    • Human review

    RunDQ eligibility engine

    The decision-tree model behind eligibility questions / Sep to Dec 2025

    The engine sometimes asked eligibility questions in the wrong order. I found out why and shipped the fix.

    Read how I tracked it down

    Finding it

    I tested all five RunDQ APIs and walked through the code with the teams who own the connected services. The cause was subtle: positive scores were split across child nodes, but negative scores weren't, so a child could end up outscoring its own parent.

    Fixing it, and after

    I proposed the fix and shipped it. I also built a conversational bot in Python that runs eligibility checks through the RunDQ API using a custom JSON mapper.

    • Python
    • Decision trees
    • API testing
    • Debugging

    Projects

    A solo hackathon build, a team project I'm leading, and two earlier builds.

    Kaagaz

    Meta Global AI Developer Hackathon, Indie Developer track / solo / Oct 2026

    Many families who qualify for a welfare scheme lose it over paperwork: a name spelled differently on two documents, an income certificate that expired, an ID card nobody brought. The rejection shows up weeks later. Kaagaz checks the papers before anyone applies. A family member sweeps a phone camera across their documents once, and Kaagaz explains what to fix by voice in Marathi, Hindi, English, Tamil, Telugu or Kannada.

    SAM 3.1 finds each document in the video and Muse Spark reads it into structured fields. A field only counts when two frames read it the same way. From there, plain code builds the household from the ration card and checks five schemes' rules, with no LLM involved. An agent then plans the fixes, but it asks the family before changing anything, and every answer in Ask Kaagaz is fact-checked by code before it's shown.

    • 32 of 32 planted document errors caught across 30 test households
    • 641 of 645 fields read correctly (99.4%) on 58 photo-like documents
    • 26 of 26 questions answered correctly and 4 of 4 trick questions refused, with no wrong answer shown
    • 756 backend tests and 32 frontend tests

    Every number comes from synthetic documents watermarked SPECIMEN. No real family's papers were used.

    Try the app Source on GitHub

    • SAM 3.1
    • Muse Spark
    • LangGraph
    • Rules engine
    • MMS-TTS
    • Modal
    • Neon Postgres
    • Netlify
    32/32 planted errors caught

    Older projects are on my GitHub

    How I work

    Forward deployed means I own the problem as well as the code. In practice it goes like this.

    1. Go where the work happens

      HD Saathi's API design came from sitting with field workers and watching where they got stuck, not from a spec.

    2. Build the whole thing

      Data model, backend, agents, frontend and deployment. Owning all of it means no gaps between the pieces.

    3. Measure before trusting

      An eval suite with an LLM judge plus hard code checks, so "it seems fine" turns into a number I can track.

    4. Ship, listen, fix

      Put it in front of real users, take their feedback seriously, and keep going after launch, like the latency work on Yojana Saathi.

    Experience

    1. Jul 2026 to now

      Associate AI Engineer (Forward Deployed Engineer)

      Haqdarshak Empowerment Solutions, Pune

      Leading HD Saathi, extending Proof Agent and continuing to improve Yojana Saathi.

    2. Aug 2025 to Jun 2026

      AI Engineer Intern

      Haqdarshak Empowerment Solutions, Pune

      Built Yojana Saathi and fixed the question-ordering bug in the RunDQ eligibility engine.

    3. Aug to Oct 2025

      Team Lead, AI Intern

      Infosys Springboard Virtual Internship 6.0

      Led five people to build the real-time sales call assistant and delivered it early.

    4. Dec 2024 to Mar 2025

      Data Analyst Intern

      ADM Education & Welfare Society, remote

      Cleaned and transformed over a million rows of sales and transaction data in Python, and built Power BI dashboards that replaced manual reporting. Named Intern of the Month.

    5. Dec 2024 to Jan 2025

      Machine Learning Intern

      Orinson Technologies, remote

      Forecast climate indicators by comparing ARIMA baselines with XGBoost and LSTM on RMSE and MAE, and built an NLP resume analyzer.

    6. 2022 to 2026

      B.Tech, Computer Science (Data Science & Machine Learning)

      Lovely Professional University

    Dhanushragav smiling, wearing a dark suit, standing in front of palm plants

    About me

    I'm from Erode in Tamil Nadu, studied computer science at Lovely Professional University in Punjab, and now work in Pune. My work sits where AI meets India's welfare system: the same scheme can involve different paperwork, languages and people depending on the state.

    I'm most interested in AI that has to survive contact with real users, real workflows and messy government data. A demo that works once isn't the goal. I want something a field worker can trust on a bad network with a family waiting.

    தமிழ்Native

    EnglishFull professional

    हिन्दीWorking

    తెలుగుLimited

    Tools I use

    AI and LLMs
    Google ADK, A2A, MCP, LangChain, LangGraph, RAG, FAISS, Vertex AI fine-tuning, Gemini, prompt caching, context compaction, LLM evals, speech-to-text
    Backend and frontend
    Python, FastAPI, Django, REST APIs, SQL, React, JavaScript, HTML, CSS, Java
    Data
    MySQL, PostgreSQL, Cloud SQL, Firestore, Pandas, NumPy, scikit-learn, Power BI
    Cloud and DevOps
    GCP (Vertex AI, Cloud Run), AWS (S3), Docker, Jenkins, GitHub Actions, Git

    Certifications

    • Google Cloud Generative AI Leader, valid until May 2029
    • IBM DevOps and Software Engineering Professional Certificate
    • Google Advanced Data Analytics Professional Certificate
    • Google 5-Day AI Agents Intensive

    Achievements

    • Rajya Puraskar, Bharat Scouts and Guides
    • 500+ LeetCode problems solved
    • 2nd place, ONE India & One World 2025

    Get in touch

    If you're hiring for applied AI or forward deployed roles, building something for India's public sector, or just want to compare notes on agent evals, I'd like to hear from you.

    er.dhanushragav2022@gmail.com LinkedIn GitHub