In progress
Learn Doubt Solver
StudentsType, speak or photograph a question and get a step-by-step solution matched to the student's board and class.
Vision readerSyllabus retrieverStep solverHint planner
Six industries · one core
Six industries. One intelligence.
Not one general model with six prompts. Each industry has its own adapter, its own sources and its own agents on the shared Yosha Core — so answers never cross wires.
- Verticals
- 6
- Capabilities
- 25
- Source corpora
- 20
6 verticals · 1 core
one core · every verticalEvery vertical gets its own adapter on one set of weights. The router picks the adapter before the first token, and retrieval stays inside that vertical's own corpus.
The industries
Six industries. One powerful core.
Pick an industry to see what it does, what it reads and the agents behind it. Only the core is shared.
The verticals
Each vertical is an adapter over the same Yosha Core weights, with its own retrieval corpus. Switching vertical switches the adapter, the corpus and the permitted tools — never the base model.
Method
Engineered, not prompted.
Every industry is built in layers: a remixed corpus, a dedicated adapter, grounded retrieval — and agents that ship.
03
Industry Adaptation
Continued pretraining plus a swappable adapter per vertical.
- Continued pretraining on regulatory and statutory text
- One LoRA adapter per industry, hot-swappable at serving time
- Router selects the adapter before the first token is generated
04
Retrieval
Indic RAG over primary, citable sources.
- Hybrid dense + lexical retrieval over statute and circular corpora
- Section-level chunking so a citation resolves exactly
- Freshness pipeline for amended rules and new advisories
05
Agent Runtime
Planning, tool calling, memory and sandboxed execution.
- Planner that decides retrieve, call a tool, or answer directly
- Typed tool registry with per-industry permissions
- Sandboxed code and document execution with audit logs
- Vision front-end reads photos, scans and handwriting into text for the core
Industry midtraining — 275B of the 2.5T run budget (11%). The five mapped industries (trade to follow) re-mixed with higher quality weighting and longer documents, so the base model already leans toward the work it will be asked to do.
Evaluation
To be measured by experts.
Held-out test sets, one per industry (trade to follow), to be built and graded with industry experts and automatic scorers. Figures are Yosha 100B targets, not results — the eval card will ship with them.
Education
Yosha LearnIndiaExam-QA
Class 6–12 and competitive exam question answering
Target78.4%
Training corpus225 B tokens
Capabilities5
Banking & Finance
Yosha FinanceIndiaBank-QA
Regulatory circular comprehension with citation accuracy
Target84.1%
Training corpus175 B tokens
Capabilities5
Law
Yosha LegalIndiaLawBench
Section identification under BNS / BNSS / BSA
Target72.9%
Training corpus150 B tokens
Capabilities5
Cybersecurity
Yosha ShieldSOCTriageBench
Alert severity agreement with analyst ground truth
Target81.7%
Training corpus100 B tokens
Capabilities5
Healthcare
Yosha HealthIndiaClinicalBench
Guideline-grounded answers on Indian clinical cases
Target76.3%
Training corpus100 B tokens
Capabilities5
Targets only. The held-out sets are still to be built; eval cards publish with Phase 03 of the roadmap.