Three systems, in enough detail to check.
We would rather show one system properly than twenty logos. CourtNetra is ours and it is the deepest: hybrid retrieval over 18,863,754 judgments, a citator that says whether an authority is still good law, and an automated evaluation that fails the build when hallucination goes above 2%. Outbrew and Savidhi are also ours, and also live.
CourtNetra
Our own legal research platform, also shipped as NyayLens. Indian case law is the hardest retrieval problem we could have picked for ourselves: the corpus is enormous, the same case is cited differently across the AIR, SCC and SCR reporter series, and a wrong answer lands on a professional who is inclined to trust it. It is the reason we can talk about grounding as a build gate rather than as an intention.
Figures from the production system, not a pilot and not a projection. Ask us for the CI report.
What is actually in there
- 18,863,754 judgments: 39,158 Supreme Court and 18,824,596 High Court, across all 25 High Courts, 1950 to the present.
- 687,289 chunks from the subset selected for vector retrieval, embedded with Qwen3-Embedding-4B at halfvec(2560), indexed under HNSW in PostgreSQL with pgvector, on a self-hosted CUDA GPU pipeline.
- Hybrid retrieval: dense pgvector similarity and tsvector BM25, reconciled by Reciprocal Rank Fusion, inside a hard 30-second budget.
- A Citator classifying every authority as good law, distinguished, doubted, partially overruled or overruled, across a 165,000-entry AIR-SCC-SCR crosswalk.
- 268 API endpoints, a 41-page Next.js 14 web app, and a 37-screen React Native app live on Google Play.
- Built and operated solo by our founder, Aniruddh Atrey.
Architecture, the trust layer, and the query planner bug that took the longest to find.
Work our founder did somewhere else
This section is separated from the one above deliberately. Everything before this point is a system MetaMinds built, owns and operates. Everything here is work Aniruddh Atrey did at a prior employer, where MetaMinds was not the contracting party and no client relationship existed. Mixing the two is how a portfolio quietly becomes a claim about clients, so they do not get mixed.
SaveLIFE Foundation road safety systems
Prior employerThree production systems built by Aniruddh Atrey while employed at SaveLIFE Foundation, before MetaMinds existed: a multi-stage Haystack v2 retrieval pipeline that cut analyst lookup time approximately 70% across 100+ reports a week, a WhatsApp learning channel built for exactly-once delivery on a 12-state machine, and a computer vision pass over 500+ hours of dashcam footage at 95% precision. MetaMinds was not the contracting party on any of them.
Including the five things these numbers do not mean.
Two more products we built and run
Both are in production. Neither has a number on this page, because we publish figures we can evidence and these are not measured to that standard yet. When they are, they will appear here with their provenance attached.
Outbrew
Ours. Live.AI cold email and outreach
An AI cold-email and outreach platform. We built it, we run it, and it is in production today.
No metrics published. Ask on a call and we will open the system and show you.
Savidhi
Ours. Live.Booking and scheduling
A platform for booking temple services and consultations. We built it, we run it, and it is in production today.
No metrics published. Ask on a call and we will open the system and show you.
Client work sits under NDA.
Most of what we build for clients is covered by a non-disclosure agreement, so it is not on this page and it will not be. That is the correct outcome: a firm that publishes your architecture without asking will publish the next one too.
Available on request
On a call we can walk through the parts that matter to your decision, with anything identifying removed:
- The architecture, and why each component was chosen over the obvious alternative.
- Evaluation results, including the numbers that were bad before they were good.
- The failure modes we hit in production, and what the fix cost.
- References, where the client has given written permission for us to arrange one.
We do not name a client, publish a logo or quote a testimonial without written permission.
Tell us what breaks if the AI gets it wrong.
That single answer tells us more than a requirements document, and it is how every one of the systems above started. Thirty minutes, straight to the engineer who would build it.
Typical reply within one business day · Engagements start at $2,500