Six systems, and what each one does
No client names, because most of this work sits under an NDA. No percentages, because nobody measured one we could publish. What each system takes in, what it does with it and what it hands back is accurate, and every one of these is a repository we can open on a call.
Defence electronics
Telling one radar apart from another
A receiver picks up pulses from several radars at once, interleaved, with nothing in the stream to say which pulse came from which. We build the pipeline that pulls them apart: burst detection from the gaps between arrivals, pulse-interval histograms, clustering on frequency and power, then a classifier that names each emitter. It is scored on false alarms and correct identifications, so the client can argue with the number.
Python · Signal processing · Classification
- Goes in
- Interleaved pulses from several radars at once
- Then
- 1Burst detection from the gaps between arrivals
- 2Pulse-interval histograms
- 3Clustering on frequency and power
- Comes out
- Each emitter named, with a confidence against it
- Held to
- Scored on false alarms and correct identifications, so the number can be argued with
Road tunnels
Four screens, one operator, one shift
The person on shift gets a live 3D model of the tunnel, the incident procedure they have to follow, the equipment panels and the camera wall. It reads the SCADA system and the PLCs directly. It also keeps the audit trail the safety case depends on, so a compliance officer can open a read-only view and see what happened and when.
Unreal Engine · Next.js · MQTT · Fastify
- Goes in
- The SCADA system, the PLCs and the camera wall, read directly
- Then
- 1A live 3D model of the tunnel
- 2The incident procedure for what is happening now
- 3Equipment panels within reach of the same screen
- Comes out
- One operator holds the whole tunnel through a shift
- Held to
- Everything written to the audit trail the safety case depends on, in a view a compliance officer can open
Clinical care
Decision support for a doctor seeing sixty patients a day
A doctor working at that rate cannot stop to type. So the speech recognition runs on the phone itself, offline, in ten Indian languages, and the clinical reasoning runs in the cloud. Every suggestion carries the published guideline it came from. The whole interaction adds under ninety seconds to a consultation, which is the constraint the design was built around.
On-device ASR · React Native · Guideline-grounded output
- Goes in
- The doctor speaks, in one of ten Indian languages
- Then
- 1Speech recognition on the handset itself, offline
- 2Clinical reasoning in the cloud
- 3Every suggestion carrying the published guideline it came from
- Comes out
- A prescription the doctor edits and signs
- Held to
- Under ninety seconds added to a consultation, which is what the design was built around
Autonomous systems
Turning plain language into fleet commands
A commander says what they want. The system turns that into playbook commands for whichever mix of air, ground and surface platforms is on the map. It refuses anything that breaks a fleet-level constraint, and it explains each decision in words a person can argue with. It sits above the existing autonomy stacks rather than inside them, so no platform has to be modified.
LLM intent translation · Vehicle-agnostic playbooks · Audit trail
- Goes in
- A commander says what they want, in plain language
- Then
- 1Intent translated into playbook commands
- 2Checked against the fleet-level constraints
- 3Anything that breaks one refused, in words a person can argue with
- Comes out
- Whatever mix of air, ground and surface platforms is on the map
- Held to
- Sits above the existing autonomy stacks rather than inside them, so no platform is modified
Recruitment
Shortlisting that leaves a paper trail
Jobs and applicants come from the client's existing tracking system. The pipeline parses every CV into a structured profile, ranks candidates against a written rubric rather than a vibe, and runs the outreach over email, WhatsApp and voice. We designed the consent, retention and subject-access handling in at the start, because retrofitting it into a recruiting database is close to impossible.
Next.js · Postgres and pgvector · GDPR and DPDP by design
- Goes in
- Jobs and applicants from the tracking system already in use
- Then
- 1Every CV parsed into a structured profile
- 2Ranked against a written rubric rather than a vibe
- 3Outreach over email, WhatsApp and voice
- Comes out
- A shortlist with the reasoning attached to each name
- Held to
- Consent, retention and subject access designed in at the start, because retrofitting them is close to impossible
Our own product
AIAuditSense
The compliance auditor we built for our own consulting work and then opened up. It reads a website, a repository or a product description, then works out what the system is under the law. It scores every dimension of three frameworks and hands back the gaps, with the clause references attached.
EU AI Act · NIST AI RMF · DPDP Act
- Goes in
- A website, a repository or a product description
- Then
- 1Work out what the system is under the law
- 2Score every dimension of three frameworks
- 3Locate the gaps and rank them
- Comes out
- A report with the clause references attached
- Held to
- Written to get you oriented. It is not legal advice and says so
Those schematics are drawn rather than screenshotted. Four of these systems run inside somebody else’s business, and their screens carry that client’s name, their staff and, in two cases, real people’s records. Drawing the mechanism instead is the part we are free to show, and it is the part you would have been looking at anyway.
Whichever of these is closest to your problem
Tell us which one it is and we will walk you through it properly, including the parts that did not work first time. Thirty minutes, no deck.