Skip to main content
Adelard D'Cunha

AI/ML engineer & Data Scientist

AI/ML Engineer

I'm an AI/ML engineer working on generative and agentic systems, multi-agent platforms that let people query their own data in plain language, RAG pipelines over documents and video, and the chat and voice layers sitting on top of them. A good share of the work is the part nobody demos: sandboxing what agents write, keeping every answer traceable to a source, and watching models for drift after they ship.

Maharashtra, IndiaOpen to new opportunities
GitHubLeetCodeKaggleStack OverflowTwitterLinkedInMediumView Resume

About

I build generative and agentic AI systems that run in production. At EduvanceAI, that's meant multi-agent platforms that let consultants query their own datasets in plain language, RAG pipelines spanning over a thousand hours of video and thousands of documents, and the chat and voice layers built on top of both.

A good share of the work is the part nobody demos: sandboxing code the agents write, keeping every answer traceable back to a source page or timestamp, and watching models for drift once they're live. I care as much about the guardrails as the model itself.

Outside of agentic and generative work, I've built a real-time card fraud detection service scoring transactions inline under 40ms, and hierarchical demand forecasting for multi-store retail: my closest brushes with BFSI and traditional MLOps respectively.

Strengths

  • Multi-agent systems & agentic orchestration
  • RAG architecture & retrieval evaluation
  • Production MLOps: drift detection, monitoring, safe rollout
  • Applied deep learning, NLP & speech

Focus areas

  • Generative AI
  • Agentic systems
  • RAG & retrieval
  • MLOps & applied ML

70

Outlets optimized

40+

Technologies used

2

Certifications

1

Publications

Skills & Expertise

Agentic & Generative AI

  • Retrieval Augmented Generation (RAG)
  • Agentic Workflows, Chatbots
  • Prompt Engineering (Chain Of Thoughts, Few Shots, Zero Shot, ReAct Prompting, Guardrails)
  • Conversational AI
  • LLM Harnessing
  • Quantization
  • Paged Attention
  • KV Caching
  • Flash Attention
  • Knowledge Distillation
  • Inference & Cost Optimization

Retrieval & Storage

  • Pinecone
  • ChromaDB
  • Neo4j knowledge graphs
  • Redis
  • PostgreSQL
  • Chunking
  • Semantic Search
  • recall@k evaluation

ML, Deep Learning & NLP

  • Supervised Learning (Regression and Classification)
  • Unsupervised Learning (Clustering, Dimensionality Reduction, Association Rules, Anomaly Detection)
  • Reinforcement Learning (Model Free, Policy Based)
  • Semi-Supervised Learning
  • Time Series Forecasting (RNN / GRU / LSTM)
  • Transformers
  • Feature Engineering
  • Model Evaluation & Testing
  • Applied Statistics
  • Probabilistic Modeling
  • Hypothesis Testing

Languages & Frameworks

  • Python
  • SQL
  • TensorFlow
  • Keras
  • scikit-learn
  • Hugging Face Transformers
  • Pandas
  • NumPy
  • Node.js

Data & Infrastructure

  • PySpark Structured Streaming
  • Apache Kafka
  • Docker
  • Kubernetes
  • AWS (EC2, SageMaker)
  • REST APIs & microservices

MLOps

  • MLflow
  • Drift detection (PSI & KS)
  • Automated retraining
  • Champion–challenger promotion
  • Shadow & canary rollout
  • Model monitoring

Where I've worked

  1. Feb 2025 to Present

    India

    Full-time

    Software Development & AI Engineer

    EduvanceAI

    • Talk2Data: built an agentic LLM system letting consultants query structured data in plain language or speech, fully automating feature engineering, EDA, and analysis; used RAG + MemPalace for dynamic, semantic context fetching to cut token use and inference cost ~95%; added a ReAct-based intent-decomposition agent that splits queries into sub-tasks and routes them correctly; designed a synchronous tool generator that raised output consistency on repeated queries from ~65% to 100%; hardened the system with a Docker sandbox and static guardrails against unsafe code; split it into independently scalable Kafka services (sandbox, intent, tools, etc.), hitting 99% accurate answers; and shipped a Next.js chatbot with live socket streaming so users can watch the agent's reasoning in real time
    • Virtual Coach: built a multimodal RAG app returning the exact video timestamp and manual page a field technician needs; chunked videos into 10-second overlapping segments with frame-level breakdown and used hybrid dense + BM25 indexing and reranking, applying the same 50-token overlapping-window approach to documents for consistent ranking; used Kafka to fan out queries and merge results from both modalities asynchronously; switched from OpenAI embeddings to a self-hosted, open-source quantized multilingual model (AkshitaS/bhasha-embed-v0) after benchmarking, dropping monthly embedding cost from $512 to ~$80; and built an audio library (VAD, noise cancellation, silence trimming, thinking/breath-pause detection) that improved throughput 40% and cut cost per minute ~60%
    • Sales Copilot: created a GenAI sales-lifecycle tool that plans monthly dealer visits, recommends pitches from past visits, suggests products, and turns post-visit audio notes into the next plan; built a journey planner solving multi-constraint TSP routing across 70 outlets in South & East India, cutting travel distance 70% while lifting coverage 50%; added admin-managed dealer targets, GenAI pitch generation from visit history, and a hybrid collaborative + content-based recommender combining time-series forecasting, dealer clustering, and TSP, plus audio-derived Next Best Actions; and built a real-time error-logging and model-monitoring service that cut downtime 99% (from constant to about once a week) and doubled the speed of root-cause resolution
    Day-to-day responsibilities
    • LMS (Learning Management System): a GenAI-powered platform for scenario simulations (e.g. manager-to-employee confrontations, sales and marketing scenarios); built role-based access control (Super Admin / Admin / User), an SMTP notification microservice, and the Eduvance.ai marketing site for a soft-skills LMS with embedded voice and chat AI
    • RAG
    • Agentic Workflows
    • Apache Kafka
    • Docker
    • Next.js
    • Hybrid Retrieval (Dense + BM25)
    • Time-Series Forecasting
    • TSP Optimization
    • RBAC

  2. March 2024 to December 2024

    Internship

    Data Analyst Intern

    Forbes Advisor

    • Text-To-SQL: built an NLP interface on PostgreSQL converting natural-language questions into SQL for financial reporting and business dashboards; pitched the POC to stakeholders and it was later adopted full-time; used an LLM layer to translate requests into SQL, fetch data, and render results in a React.js app; added JWT-based auth to scope data access by user role/department; and implemented SQL validation and guardrails against unsafe queries and prompt injection, lifting data-fetch accuracy to 99%
    • ETL & Data Study: built web-scraping pipelines with Scrapy and Selenium to collect, clean, normalize, and stage external financial data for predictive modeling; scheduled weekly, monthly, and yearly extraction jobs with Apache Airflow; built automated data-quality checks for both scraped and database data, loading valid records into MongoDB and quarantining/alerting on failures
    • PostgreSQL
    • LLM (Text-to-SQL)
    • React.js
    • JWT Auth
    • Scrapy
    • Selenium
    • Apache Airflow
    • Databricks
    • Excel
    • MongoDB

  3. Janurary 2024 to March 2024

    Remote (Bangalore)

    Internship

    Data Analyst Intern

    Catalyst Management Services

    Day-to-day responsibilities
    • Increased data collection efficiency using Beautiful Soup, Parser and Selenium to streamline scraping processes from diverse sources by performing mass web scraping on over 250+ websites as data sources.
    • Developed automated ETL pipelines, reducing data update processing time and enhancing data availability for analysis.
    • Conducted rigorous and multiple levels/phases of data cleaning and validation with Pandas, ensuring dataset accuracy and reliability through well-documented protocols. Automated the whole process into a single system by the end of the sprint.
    • Python 3.9
    • node.js
    • Docker
    • ngrok
    • PostgreSQL
    • LLM (Text-to-SQL)
    • React.js
    • Scrapy
    • Selenium
    • Bert-uncased
    • Transformer
    • Self-Attention
    • MongoDB

  4. March 2023 to December 2023

    Remote (Navi Mumbai)

    Internship

    Data Scientist Intern

    Plan My Venture Pvt. Ltd.

    • Text-To-SQL: built an NLP interface on PostgreSQL converting natural-language questions into SQL for financial reporting and business dashboards; pitched the POC to stakeholders and it was later adopted full-time; used an LLM layer to translate requests into SQL, fetch data, and render results in a React.js app; added JWT-based auth to scope data access by user role/department; and implemented SQL validation and guardrails against unsafe queries and prompt injection, lifting data-fetch accuracy to 99%
    Day-to-day responsibilities
    • Developed predictive models achieving 85%+ accuracy using advanced statistical methods and hypothesis testing; utilized rigorous cross-validation and evaluation metrics to drive capital allocation decisions.
    • Fine-tuned BERT-based transformers on domain-specific financial text to extract latent sentiment and predictive insights, significantly outperforming baseline models in unstructured data classification.
    • Conducted in-depth EDA to isolate key business drivers and correlations, directly informing feature selection and hyperparameter optimization for production-scale forecasting.
    • Built automated data pipelines integrating Shopify-based analytics with custom dashboards, enabling near-real-time monitoring of revenue and retention KPIs.
    • Python 3.9
    • node.js
    • Docker
    • ngrok
    • PostgreSQL
    • LLM (Text-to-SQL)
    • React.js
    • Scrapy
    • Selenium
    • Bert-uncased
    • Transformer
    • Self-Attention
    • MongoDB

Academic background

  1. 2021 — 2024

    Fr. Conceicao Rodrigues College of Engineering

    Bachelor of Engineering, Artificial Intelligence and Data Science

    Mumbai, India

  2. 2018 — 2021

    St. Xavier's Technical Institute

    Diploma, Electronics and Telecommunication

    Mumbai, India

Certification & Credentials

  • Machine Learning Specialization

    DeepLearning.AI & Stanford University (Coursera)

  • Mathematics for Machine Learning and Data Science

    DeepLearning.AI (Coursera)

Writing & research

  1. Your ASR Model Isn't the Problem, Your Signal Path Is

    Medium · Adelard D'Cunha

    On why audio signal quality: not the ASR model itself: is usually the real bottleneck in speech pipelines.

    Read more →

A little more personal

Work is one part of the picture. Here's a little of what fills the rest of it.

Outside of work

Long-distance cycling and slow, deliberate cooking.

Cuisine

Coastal Konkani food, and a long-running fondness for good masala tea.

Sport

Love for badminton or football stays constant.

Reading

Currently reading `100 ways to survive in the wild`

Travel

Prefers slow travel: a week in one place over five cities in five days.

Let's Connect

Interested in building something meaningful together?

I'm always glad to hear about interesting problems, thoughtful teams, and roles worth considering.

adelarddcunha07@gmail.com