01
Sovereign fine-tuning
Adapting models to one organization’s language, data and tasks, deployed in its own environment.
Oshara / Research
We work below the API.
Our R&D team fine-tunes and evaluates models for individual organizations, and engineers the harness they run in. Here is a selection of that work.
01Research areas
What we research.
02
Harness engineering
Orchestration, tool use, guardrails and human approvals that keep agents predictable in production.
03
Evaluation
Task-specific test suites that measure what matters to the business, before and after every change.
02Projects
Client and in-house work.
We build and evaluate production-oriented AI systems for client organizations, spanning model fine-tuning, agentic workflows, document intelligence and evaluation infrastructure.
Fine-tuning
Emotion-aware speech generation
Fine-tuned Oshara’s text-to-speech models for expressive, emotion-aware storytelling, so generated speech conveys different emotional tones through explicit emotion tags for dialogue, narration and character-driven stories.
Result
Emotion-conditioned speech via explicit tags, extending Oshara TTS beyond neutral speech toward expressive storytelling
Agentic systems · Client engineering
Workflow automation and document intelligence
Worked with Zakipoint to automate operational workflows and build a document parsing engine, designed around the organization’s existing architecture and operational requirements, with supporting work on QA agents.
Result
A document parsing solution built around the client’s architecture and agentic automation workflows supporting broader enterprise AI adoption
Evaluation
Agentic AI evaluation framework
Developed evaluation approaches for agentic systems: assessing agent behavior, analyzing outcomes and comparing performance across implementations and iterations, going beyond simple output checking to judge the quality of agent-driven workflows.
Result
A foundation for systematic agent evaluation and informed comparisons during development
HarnessOngoing
Agentic harness and workflow orchestration
Building an agentic harness for developing, running and evaluating agent-driven workflows: structuring how agents operate within a workflow, coordinating their execution, and observing and assessing their behavior.
Result
A reusable foundation for developing, orchestrating and evaluating agentic AI systems
03Papers and writing
What we've learned, written down.
Research outputs and practical learnings from building, fine-tuning and evaluating AI systems.
Paper · 2026
ForthcomingEmotion-Conditioned Speech Generation with Oshara
Research on emotion-aware TTS fine-tuning and expressive speech generation. Manuscript and publication details pending.
Talks and open source
All posts on the blog- Sep 7, 2026Teaching XTTS v2 to Feel: English First, Then NepaliSeven emotion tags for XTTS v2, first in English, then in Nepali, using a dataset built from Nepali films.
- Jul 28, 2026Fine-Tuning IndicConformer for Nepali Speech RecognitionBenchmarking IndicConformer against six Nepali speech-to-text systems, then fine-tuning it and fixing a tokenizer bug.
- Jul 28, 2026Two LLMs Are Better Than One: How We Build Low-Latency Voice AgentsA fast model for responsiveness alongside a smarter model for the answer, inside a real-time voice pipeline.
- Jul 22, 2026A Nepali Voice for XTTS v2Fine-tuning XTTS v2 for Nepali by reusing its Hindi tokenizer: the best open, self-hostable Nepali voice we could measure.
- Jul 17, 2026Building an Agent, Step by StepCreating a voice agent on Oshara.ai, from the My Agents page to an embeddable widget.
- Jul 17, 2026Teaching a Duplex Voice Model to Call ToolsAdding a tool-calling action stream to PersonaPLEX, a full-duplex audio model, over four training attempts.
- Jul 17, 2026Teaching PersonaPLEX to Recite ShakespeareLoRA and spoken question-answer data from a Shakespeare play, teaching PersonaPLEX to recite grounded lines aloud.
- Jul 1, 2026Teaching XTTS v2 to Speak NepaliA Nepali narrator from XTTS v2: why epoch 10 beat epoch 30, and the text-pipeline fixes that stopped babbling.
- Jun 19, 2026Teaching a Machine to Speak NepaliFine-tuning Chatterbox v3 for Nepali: tokenizer bugs, phantom sentences and an overfitting hunt.
The team
A small R&D team of engineers working across speech and agents. We fine-tune text-to-speech and speech recognition models, including for low-resource languages like Nepali, build the harnesses and evaluations that agentic systems run on, and deliver that work into client organizations.
A model fine-tuned for your organization, that belongs to you.
Talk to our R&D team about sovereign fine-tuning, harness engineering and evaluation, or hear our voice work in action.