AI Engineer, based in Bengaluru. Eight years in data and AI engineering, the last three building AI products end to end, seven of them fully remote with teams in Hong Kong, Europe and India. I build tools that turn messy information into data you can actually use — reading web pages and documents, pulling out what matters, and checking the result is right before anyone relies on it.

Master's in Data Science & AI, IISER Tirupati, 2026 (CGPA 8.9/10). English, Hindi, Bengali, and working Nepali.

Looking for AI engineering work — Bengaluru or remote. Available immediately. irfan.ali@datacortex.in

Full writeups at datacortex.in. Every number below comes from a published source you can check.

Finds 87% of the products a company actually sells, for about 5 cents a company.

Reads a company's website and returns the products they sell as organised data, not plain text. It works in five steps: search, read the pages, list candidate products, merge duplicates, then decide which are real. When it isn't sure, it holds the answer back for a person to check.

• Scored, not eyeballed. I built the test set: companies whose correct answers I checked by hand, kept fixed so every change is scored against the same list. Nothing shipped unless it held up on that set.
• Anyone can re-check the score. The labels, frozen predictions and scoring script are public. python3 rescore.py recomputes the numbers with no API keys.
• Cheap to run. The full 20-company evaluation cost $1.10 in model and search calls, and no company was left out of the score.

Case study · Test set and scoring script

FastAPI DSPy Pydantic LangSmith OpenAI

Built so an AI provider outage never loses a call.

Handles weekly check-in phone calls and writes up what was said. Running in production for evaluation.

• Safety before any model. Safety checks run on the raw transcript first and flag anything as critical or elevated.
• Three layers of fallback. If one analysis provider fails it moves to the next, then to fixed rules, instead of returning an error.
• Each call saved exactly once, even when the phone provider reports the same call twice.
• Memory without giant prompts. Context from earlier calls comes from searching past transcripts, not stuffing everything into one long prompt.

Case study

Next.js Bolna Groq pgvector

Co-founded and built solo. A live product that replaced four nightly notebooks with one set of books, in daily use by its first distributor.

"Before Godam, I balanced four notebooks every night — cash, godown, cheques, udhaaro." — Sumit Pokhrel, co-founder and first user, Suju Construction & Trading

Stock, sales, cash, cheques, credit and field visits for distributors who supply 40–80 retailers across a district, built for a phone screen.

• Shaped by daily use. Built in April 2026, then four rounds of features shipped from the distributor's real requests: receipts and vouchers, bill-by-bill payment matching, cheque reminders, Tally-style statements.
• Speaks the business's language. Statements in Nepali (Bikram Sambat) dates, VAT and PAN on every party, sauda (advance deal) book, credit aging, and one-tap PDF sharing on WhatsApp.
• Built for the field. Installs on any phone from the browser, because daily sales and collections are entered on the road, not at a desk.
• Permissions that protect the books. Owner, staff and view-only roles; view-only users can never change the accounts.
• No AI in the critical path. Getting the numbers right comes down to careful data handling and permissions.

Product site · Case study

Next.js TypeScript Supabase Postgres PWA

Python libraries I build and maintain on PyPI. Numbers below refresh daily.

Library	What it does	Live stats
toxic-comment-classifier	Detects toxic comments, with a score per category (toxic, insult, threat and more)
ragfallback	Stops a search-and-answer system failing quietly. Rewrites weak queries, scores how confident the results are, and falls back when they're poor
lingo-nlp-toolkit	Text-processing toolkit on Hugging Face models: classification, entities, sentiment, question answering, summarisation, embeddings
nepal-gov-agent	Answers questions about Nepal government policy and legal documents, in Nepali and English, citing the sentence each answer came from
AgentEnsemble	Running several AI agents together: routing work between them, planning, and tracking what it costs. Works with OpenAI or Mistral
AgentCare	Voice AI for healthcare: call intake, pulling out the details that matter, chasing missing information, booking appointments
RAGNav	Search that combines keyword matching with meaning-based matching, so a question can be found either way. Answers cite the passage they came from
AskPandas	Ask questions about a CSV in plain English. Runs on a local model by default, so nothing leaves your computer, or on OpenAI / Mistral
PyroChain	Using AI agents and a fine-tuned language model to work out which features matter in a dataset

All packages: pypi.org/user/irfanalidv

Sole-author, peer-reviewed (double-anonymised) and open access, written at IISER Tirupati. Research interests: natural language processing, information extraction, conversational AI, neural information retrieval.

Indian Journal of Artificial Intelligence and Neural Networking, vol. 6, no. 1, pp. 10–17, Dec 2025 · sole author · IISER Tirupati

Benchmarks conversation AI for mental-health support on 16,084 therapy conversations (6,338 real clinical interviews). A BERT classifier reaches 86.7% accuracy and 86.1% F1, and the gaps between models are checked with McNemar's test and bootstrap confidence intervals (p < 0.05) rather than taken at face value. The paper also documents the dataset's limits, including no demographic labels, and states that clinical validation is needed before any real-world use.

_{Keywords: Mental health · Conversational AI · BERT · Neural networks · Sentiment analysis}

Indian Journal of Artificial Intelligence and Neural Networking, vol. 5, no. 6, pp. 10–19, Oct 2025 · sole author · IISER Tirupati

Tracks what restaurant customers talk about, and how it shifts over time, across 3,000 Yelp reviews from 1,467 businesses in 248 metro areas (2005–2018). Combines transformer-based extraction (97.2% F1 on named entities), topic modelling (eight topics, 87.5% diversity), sentiment forecasting (17.9% mean absolute error) and city-level comparisons that are statistically tested (p < 0.05 across 10 cities).

_{Keywords: Topic modelling · Information extraction · Neural-symbolic learning · Temporal analysis · Business intelligence}

Journal: Indian Journal of Artificial Intelligence and Neural Networking (IJAINN), Lattice Science Publication · ISSN 2582-7626 (online) · indexed in Crossref

irfan.ali@datacortex.in · LinkedIn · datacortex.in