Area 01
Applied machine learning
Models that survive contact with reality — trained on messy, real-world data and held to operational, not benchmark, standards.
Resnova Research
We build the systems, then we study how they behave.
Resnova Research is the lab behind the products — applied work on machine learning, computer vision, low-resource language, and the hard question underneath all of it: how to deploy autonomous systems that institutions can actually trust.
0
Active research areas
0+
Papers & technical reports
100%
Findings from real deployments
0
Open problems we are funding
What we study
Six areas. One standard: it has to work in the field.
Resnova Research is applied, not speculative. Every line of inquiry traces back to a system running under real load for a real institution.
Area 02
Computer vision at the edge
Vision systems that infer on-device, retain decisions rather than footage, and run in single-digit watts at the perimeter.
Area 03
Language & low-resource NLP
Bangla-first natural language — understanding, retrieval, and generation for a language the global models treat as an afterthought.
Area 04
Edge & embedded intelligence
Inference on microcontrollers and SBCs — Jetson, Raspberry Pi, ESP32 — that degrades gracefully when the network does not exist.
Area 05
Sovereign & auditable AI
How to deploy autonomous systems in regulated settings: provenance, attribution, key custody, and the right to audit every decision.
Area 06
Systems & infrastructure
The unglamorous substrate — data pipelines, orchestration, and reproducible deployment that lets a model run on a Tuesday under real load.
Featured paper · 2026
Decision ledgers: audit-grade provenance for autonomous systems in regulated settings
We formalize the "decision ledger" — an append-only, cryptographically chained record of every consequential action an autonomous system takes — and argue it is a precondition, not an afterthought, for deploying AI in government and critical infrastructure.
Publications
Papers, reports, and field notes.
We publish what we learn in deployment — including what we got wrong. Everything here is grounded in a system we actually shipped.
2026
Working paper · Sovereignty
Decision ledgers: audit-grade provenance for autonomous systems
S. S. Adnan · Resnova Engineering
Read →
2026
Technical report · Edge AI
Sub-100ms number-plate recognition on a $99 board: a deployment study
Resnova Engineering
Read →
2026
Field note · Fintech
Instrumenting the informal economy: identity before credit
S. S. Adnan
Read →
2025
Working paper · NLP
Bangla-first retrieval: closing the low-resource gap for civic search
Resnova Engineering
Read →
2025
Technical report · Vision
Retain the decision, not the frame: privacy-preserving civic vision
Resnova Engineering
Read →
2025
Field note · Systems
Reproducible sovereign deployment on the National Data Centre
Resnova Engineering
Read →
2025
Working paper · Sovereignty
Four properties of a system a state can actually own
S. S. Adnan
Read →
2024
Technical report · Edge AI
Bridging RS-485 to MQTT: forty years of industrial telemetry, one gateway
Resnova Engineering
Read →
Open problems
The questions we have not answered yet.
Resnova publishes its open problems. If you work on these, we want to hear from you.
Open problem 01
Verifiable inference
How does a citizen verify that the model which decided their case ran as audited — without trusting the operator? We are exploring cryptographic attestation of inference.
Open problem 02
Bangla at the frontier
Frontier models still treat Bangla as low-resource. What does a genuinely Bangla-first model — not a translated one — require in data, tokenization, and evaluation?
Open problem 03
Graceful sovereignty
Sovereign infrastructure is slower to scale than hyperscale. Can we get the accountability of in-country hosting without the elasticity penalty? We think hybrid custody is the key.
Work with the lab
Research that ships, or it isn't research.
We collaborate with institutions, universities, and engineers who want their work to leave the page. If that is you, start a conversation.