Voice and conversation products, built and launched in three weeks.
A pod of senior forward-deployed engineers who build products on speech-to-text and language models: transcription, conversation analytics, call automation and meeting assistants. In the languages your users speak, and on your infrastructure if you need it.
What the pod builds
Pick the product
The pod builds it and launches it.
Speech-to-text products
Transcription pipelines and the products on top of them: recordings and live calls, the accents and languages your users speak, your terminology.
Conversation analytics
Every call, meeting and chat analysed: quality and compliance checks, themes, complaints and requests — what customers actually said, counted.
Call automation
Voice agents that answer and route calls, automated follow-ups, and summaries and next steps written straight into your CRM.
Meeting assistants
Notetakers that join calls and return summaries, decisions and owner-assigned action items — like Khulasa.
Conversation archives
Searchable transcripts with retention rules, access roles and audit logs, kept inside your own infrastructure.
Built for real conversations
Real calls are messy: accents, background noise, people switching languages, product names a generic model has never heard. The pod builds for that:
- Any language your users speak, including conversations that switch between languages.
- Your product names and terminology, recognised across every recording.
- Where conversations happen: telephony, Zoom, Teams, Meet, chat and support tools.
- Speech and language models picked for your case — cloud APIs, or open models on your own hardware.
The showcase
Proof: Khulasa
Khulasa is one product this pod built — an AI notetaker we run ourselves. It’s here as proof of how we work, not as the limit of what the pod builds.
An AI notetaker in production
Joins Google Meet, Zoom and Microsoft Teams calls, records, transcribes Arabic, English and mixed speech, and produces a summary, decisions and owner-assigned action items — with Jira export, organisation roles and per-recording sharing.
Runs on-premises or in a private cloud
The speech and language models can run on the customer's own hardware; the on-premise edition runs with no outbound connection at all. Meeting content is never used to train models.
We publish the benchmark
On real code-switched Arabic–English meeting audio, the self-hosted configuration kept more than three times as many Arabic–English switch points as the comparison configurations and scored +16 to +35 points higher on English words inside Arabic speech. On overall word error rate it is at parity with leading vendors — we say so, and we show the full method.
- 3×+
- more Arabic–English switch points kept than the comparison setups
- +16 to +35
- points higher on English words inside Arabic speech
- Parity
- with leading vendors on overall word error rate — we say so
What this proves for your product: we can take a voice product from idea to production, deploy it inside a regulated environment, connect it to the tools people already use, and measure it honestly.
Deliverables
What you get in three weeks
- 01
A launched prototype
Your voice or conversation product, live with its first users.
- 02
Connected to where conversations happen
Your call platform, meeting tool, telephony or chat.
- 03
Output where your team works
CRM, Jira, ticketing, data warehouse.
- 04
Deployment you choose
Cloud, private cloud in your region, or on-premises with no outbound connection.
- 05
All code, prompts and documentation
Yours, with a plan and a fixed estimate for the next stage.
Fifteen working days
How the sprint runs
| Stage | Days | What happens |
|---|---|---|
| Scoping call | — | 30 minutes: the product, where conversations happen, who uses it. We confirm the fit and quote one fixed price. |
| Shape | 1–3 | Flows and architecture agreed. No recordings to share yet? We start on sample or synthetic audio and switch to yours when it’s ready. |
| Build | 4–10 | The product built end to end, first outputs reviewed with your team daily. |
| Launch | 11–15 | Deployed and live with first users. Handover and a plan for what’s next. |
Scoping call
- Days
- —
- What happens
- 30 minutes: the product, where conversations happen, who uses it. We confirm the fit and quote one fixed price.
Shape
- Days
- 1–3
- What happens
- Flows and architecture agreed. No recordings to share yet? We start on sample or synthetic audio and switch to yours when it’s ready.
Build
- Days
- 4–10
- What happens
- The product built end to end, first outputs reviewed with your team daily.
Launch
- Days
- 11–15
- What happens
- Deployed and live with first users. Handover and a plan for what’s next.
Deployment options
Cloud
Fastest start; data processed in the region you choose.
Private cloud in your region
UAE, Qatar, EU; your account, our deployment.
On-premises
Models on your own hardware, no outbound connection, for regulated and sovereignty-bound teams.
GDPR, UAE PDPL and Qatar PDPPL readiness; NDA by default; details on the Compliance page.
The pod
Named engineers for the whole sprint
Daily 15-minute check-in, one review per week, everything written down.
Speech and language engineering
Speech-to-text and language models, conversation analysis, quality tuning. The team behind Khulasa.
Product and integration engineering
The app your users see, your call platform, CRM and ticketing APIs, deployment on-premises or in your cloud.
Pricing
One fixed price per sprint, quoted in writing after the scoping call.
Afterwards, if you want us to take it to production: an hourly rate or a monthly retainer, quoted with the production plan. Or your team takes it — the handover is built for that.
Questions
Yes. The sprint starts with the recording layer — call platform integration or a meeting bot — and still ends with the product live on your first real conversations.
No. Khulasa is one product this pod built. The pod builds any product on speech and conversations — analytics, call automation, voice agents, transcription.
The ones your users speak, including mixed-language conversations. We try speech models on samples of your audio early in the sprint and pick what works best.
No. We can start on sample or synthetic audio and switch to your recordings when they’re ready.
If it produces audio files or a stream, yes — that covers Avaya, Genesys, Cisco, Asterisk-based PBXs and cloud providers like Twilio. Name yours on the scoping call.
Yes — the on-premises option runs the models on your hardware with no outbound connection, the same way Khulasa’s on-premise edition does.
No. Not your audio, not your transcripts. It is in the contract.
Yours. Code, prompts, configuration and documentation are handed over. We keep only our internal toolkit.
Start with your idea
Bring the idea, and a few example conversations if you have them. We’ll tell you in the call whether this pod is a fit.