विद्या · Applied AI · Bengaluru & Mumbai · Est. 2021
Vidya.ai is an applied AI engineering firm. We design, build, and ship machine learning systems for Indian enterprises — the unglamorous 80 percent that has to actually work in production, under RBI audit, across Bharat, at scale.
§01
Three sectors where Indian businesses have problems the global playbook doesn’t solve. We’ve shipped in all three.
UPI, lending, fraud, credit — under RBI scrutiny, at Indian scale.
Demand across a fragmented, festive-driven, Bharat-wide market.
Vision on the line, and a supply chain that runs 24/7.
§02
Six disciplines, each owned by a principal engineer who has shipped it in production. We don’t subcontract the hard part.
Demand, supply, pricing, and routing models that survive contact with real noise and real planners — including a festive calendar that breaks every textbook example.
Prophet · Temporal Fusion Transformers · Gurobi · OR-Tools
KYC, loan files, invoices, GST returns, and contracts — across English, Hindi, and regional languages, and across scan quality from crisp PDFs to a phone photo of a fax.
LayoutLM · ColPali · multilingual OCR · human-in-the-loop review
Inspection, measurement, and safety on the line and in the field — including the edge boxes that run it when the network doesn’t.
YOLO · SAM · TensorRT · custom edge deployment
RAG and agent systems in 10 Indian languages that cite their sources and know what they don’t know — because your customer in Tier-3 doesn’t fill forms in English.
vLLM · LangGraph · cross-encoder reranking · Indic NLP
The 80 percent: monitoring, retraining, drift detection, rollback, and the on-call rota that keeps it alive at 2am — on infra your team actually runs.
MLflow · Evidently · Kubernetes · Ray Serve
Where to start, what to build versus buy, and — more often — what not to do yet. DPDP- and RBI-aware by default.
discovery workshops · build/buy analysis · compliance review
§03
Three platforms we built once and now deploy inside client environments. Each ships with an eval harness, a runbook, and the option to take it over.
Document intelligence
Turn loan files, KYC packs, and GST returns into queryable ground truth.
Ledger ingests multilingual document sets, extracts structured fields with cited confidence, and routes uncertain extractions to a reviewer. Built for BFSI and regulated archives where every field has to be defensible to an auditor.
Deployed at 2 NBFCs, 1 insurer · 8.4M docs processed
Forecasting engine
Operational forecasts that account for the festive calendar your plan didn’t.
Acast fits hierarchical forecasts with uncertainty bands, reconciles them up and down your product and pin-code tree, and flags when the world has drifted far enough to retrain. Replaces the spreadsheet nobody trusts anymore.
In production at a national D2C brand & a kirana platform
Fraud & compliance
Catch what the rules engine misses — with reasons attached.
Beacon watches UPI, lending, and transaction streams for fraud and compliance breaches, surfaces each alert with a human-readable rationale and evidence, and learns from reviewer dispositions. No model touches a decision alone.
Live at a fintech & a bank · ₹6.4 Cr fraud caught in year one
§04
Four engagements, anonymised under NDA, with the numbers our clients let us publish. Every entry below is a system running in production today.
Fraud the rules engine missed on a UPI lending app
The problem
A UPI-lending app was approving loans in under 60 seconds — and burning 4.1 percent of disbursals on first-default fraud the rules engine couldn’t see. Manual review at that latency was impossible.
What we built
Beacon scored every application in 1.8s with a reason and an evidence pack, queued only the borderline decile for review, and learned from dispositions weekly. The model never approves alone; every decline ships with its reasons for the audit trail.
Festive-season demand across 14,000 pin-codes
The problem
A D2C brand was planning 18,000 SKUs across 14,000 pin-codes on a model that treated Diwali like any other week. Festive stockouts ran 14 percent on hero SKUs; post-festive excess tied up ₹38 Cr in working capital.
What we built
Acast rebuilt the forecast with a festive-and-regional calendar, reconciled down to pin-code-SKU-week, and wired retraining into the promotion cycle. It now runs nightly and feeds replenishment directly — no spreadsheet in the loop.
Vision on the line, and ETAs that don’t lie
The problem
Final inspection on five lines ran on human inspectors at 2am. Defect escape was 1.8 percent; a line stop costs ₹9.5 L/hour. On the supply side, supplier ETAs were guesses that broke the schedule weekly.
What we built
A defect-detection model on edge boxes at each line (inspection never depends on the network), plus a supplier-delay model trained on three years of PO and logistics data. Inspectors now adjudicate flagged frames; planners get a real ETA with a confidence band.
Credit scoring for new-to-credit borrowers
The problem
The bank was declining 61 percent of new-to-credit and thin-file applicants outright — because the bureau-centric model had nothing to score them on. Lending to Bharat was the strategy; the model wouldn’t let it happen.
What we built
A supplementary scoring model that uses permitted transaction, GST, and behavioural signals alongside bureau data, with reason codes for every decision and a drift monitor. Approved applicants are graded, not binary-passed, so the bank can price risk instead of refusing it.
§05
Five phases. You see working software by week six, not a demo at the end. Every phase exits on a written artifact you keep.
We read your data before we pitch your solution. Exit: a written findings memo, a build/buy recommendation, and a scoped first engagement — or an honest recommendation not to engage.
Deliverables: findings memo · data audit · build/buy memo · scoped SOW
We design the system before we touch a model: data contracts, interfaces, failure modes, the eval harness, and the metrics that will define “done” — with RBI/DPDP constraints designed in, not bolted on.
Deliverables: system design · eval harness · interface contracts · risk register
Iterative, in your environment, against your eval. You see working software every two weeks, not a demo at the end.
Deliverables: working software · eval reports · biweekly demos
The part nobody puts in the deck: load, drift, rollback, observability, on-call, and the runbook. We make it survive.
Deliverables: load test · drift monitors · runbook · on-call rota
We run it with you until your team can run it without us. Retraining cadence, incident response, and a quarterly health report against the original metrics.
Deliverables: quarterly health report · retraining cadence · handover plan
§06
Quotes our clients let us attribute, with the engagement and date attached. No star ratings.
Vidya.ai delivered working software that passed our RBI audit, not a slide deck about what might be possible. They wrote the findings memo in week three that our own team had been asking for since 2019.
We’d been sold “AI” three times before. Vidya.ai was the first engagement that started with our data and ended with a system our own engineers own.
They told us not to build one of the three things we asked for. That memo saved us a year and eight figures. We hired them for the other two.
§07
The engineers who own each engagement. Every Vidya.ai project has a principal on it from discovery through handover.
Co-founder & Principal Engineer
ex-Flipkart ML platform (2016–2021) · PhD ML, IISc Bengaluru
Co-founder & Head of Engineering
ex-Razorpay staff MLE (2015–2021)
Principal, MLOps
ex-Swiggy platform lead (2016–2022)
Principal, Vision & Edge
ex-ISRO computer vision, satellite imaging (2014–2022)
Principal, Document Intelligence
ex-Zoho NLP lead (2015–2022)
Principal, BFSI & Compliance
ex-HDFC risk analytics (2013–2022)
§08
No sales call. The first conversation is with a principal engineer, lasts thirty minutes, and ends with a yes, a no, or a “send us the data first.”
Bengaluru
Prestige Atrium, 4th Floor
Whitefield, Bengaluru 560066
Mumbai
301, Peninsula Park, A Wing
Lower Parel, Mumbai 400013