Engineering services · India
Corvux Mindware is an engineering services company. Teams bring us the part of the system nobody on staff has shipped before - usually a part that spans several disciplines at once, because real systems don't respect the boundaries between them.
Face-recognition pipelines running in production. Speech models served on GPU clusters. Multi-agent LLM services. Extraction pipelines measured at 95–98% accuracy. Consent and audit systems under India's DPDP Act. Native apps signed and notarised for four platforms.
01Services
Each of these is backed by systems we have actually put into production. If a line below sounds abstract, ask us for the specifics - we would rather talk about thresholds and failure modes than adjectives.
Detection, recognition and embedding pipelines that survive contact with real input - not notebooks. We tune the thresholds, score the quality of every frame before trusting it, and build the human review path for the cases the model gets wrong.
Multi-agent LLM services with typed state and reducers that let agents write in parallel without clobbering each other. Gatekeepers that skip expensive work early, batch validation that collapses N calls into one, and truth-checking so the model cannot invent what it did not find.
The whole application, front to back. React and Vite at the top, Node or FastAPI in the middle, Postgres or MySQL underneath - and native desktop and mobile when the product has to leave the browser.
OCR and LLM extraction pipelines that emit typed records and route their own failures to review instead of failing quietly. Vector similarity, full-text ranking and graph algorithms where the relationships are the point.
Terraform-defined clusters that bring GPU nodes up on demand and take them down after, plus the unglamorous pipeline that gets a binary onto a machine. We have chased production 502s down to a missing UDP rule in a network ACL.
Applied cryptography assembled from primitives rather than bolted on, and the regulatory work that has to sit on top of it - consent, audit trails and role-based access under India’s DPDP Act 2023.
02Work
Client names are withheld under NDA. The numbers are not - every figure below is measured from the delivered system, and we will walk you through any of them on a call.
Education · multi-campus
RetinaFace finds the faces, ArcFace turns each one into a 512-dimension vector, and pgvector searches it against only the students enrolled in that class. Below 0.45 cosine similarity nothing is guessed - the face goes to a teacher review queue with its five nearest candidates. Every crop is scored on detection confidence, pixel area, Laplacian blur and brightness before we trust it, and video enrolment clusters the same person across frames to keep only their sharpest crop.
Careers & upskilling
LangGraph state machines with typed state and reducers, so a resume can be enhanced by five agents in parallel and still merge cleanly. A truth-validation pass runs afterwards so the model cannot award the candidate an achievement they never claimed. A separate six-agent graph runs adaptive technical interviews. Speech is served from NVIDIA Triton - Parakeet TDT for transcription, Kokoro for synthesis - behind a gRPC gateway, with Gemini on Vertex AI and Azure OpenAI as failover.
GPU infrastructure
Terraform-defined EKS with Karpenter provisioning GPU capacity on demand and taking it away when the queue drains. Helm-packaged services, the NVIDIA GPU Operator handling drivers and device plugins, and Prometheus scraping Triton so queue depth per model is something we can see rather than infer. An idle GPU is the most expensive line in the account, so the pools are spot-first and scale to zero.
Health-insurance benefits
An eligibility-extraction worker: OCR to full-document Markdown, then schema-forced LLM passes that return validated typed records instead of JSON scraped out of prose. Payer rules, then a verify pass. Anything the pipeline cannot stand behind is flagged degraded and routed to review rather than written silently. Golden regression tests fail the build if a single field moves. The residual error is OCR dropping table cells - not the model inventing values.
Enterprise data governance · India
A multi-tenant platform for India’s DPDP Act 2023 on Fastify and PostgreSQL 16, every row scoped by tenant. The audit trail is hash-chained, so altering a past event breaks the chain - and it sits behind an interface contract, which let the client’s blockchain team put Hyperledger Fabric underneath it without touching our code. Argon2id hashing, JWT with refresh rotation and session revocation, HMAC-signed webhooks that retry and escalate after three failures.
03Stack
55 dependencies across six disciplines. The first line of each row is what we reach for by default and have shipped on more than once. The second is in production somewhere, but is not our first pick - and we would rather tell you which is which.
ML & vision
Applied AI
Product
Data
Infrastructure
Security
04Our own apps
Two products we own outright. They are not the business - they are how we keep our own tooling sharp and prove we can carry something all the way from an empty repo to a signed installer.
CorvuxLocal-first knowledge appPrivate betaMarkdown notes on your own disk with a graph database underneath and an agent that pauses on a diff before it writes. Ships a local MCP server so Claude, Cursor and Codex read the same notes.
PorttyRemote terminal · Open sourcePre-releaseA Rust daemon owns real PTYs on your machine; your phone attaches over an encrypted peer-to-peer pipe with no SSH keys, no VPN and no open ports. Detach and reattach like tmux.
05Practice
Not values. Working rules - each one costs us time or margin on every project, and we think each one is worth paying for.
01
Our extraction pipelines report accuracy against source documents. Our recognition thresholds are numbers we tuned, not defaults we inherited. When we say something works, there is a figure behind it and a test that fails if it stops being true.
02
Every pipeline we build has a defined behaviour for the input it cannot handle - a review queue, a degraded flag, a fallback. Systems that quietly drop the cases they find hard are the ones that erode trust six months after launch.
03
Signing certificates, notarization tickets, Helm overlays, secret rotation, GPU taints and tolerations, store submissions. We have taken products through Developer ID, MSIX, Play signing and TestFlight. Software that will not deploy is software that does not exist.
04
A gatekeeper that skips a low-match job before it reaches five agents. One batched validation call instead of ten. Spot-first GPU pools that scale to zero. We treat the bill as part of the architecture, because at scale it is.
05
One encrypted transport crate serves both of our own products. Shared code that has survived two independent sets of requirements is code you can build on - and it is why the second product cost a fraction of the first.
06
Everything we cut carries a reason and a trigger to revisit it. You get a live account of what is done, what was dropped, and what is waiting on a decision that belongs to you rather than to us.
06Contact
The engagements that go well for us start with a real constraint - a model that has to hold accuracy on messy input, a cluster that has to scale GPUs without burning the budget, a platform that has to satisfy a regulator. Bring us that part.
