Product 1
Salient 1.0
The AI that always shows its proof
Every AI company is racing to build a bigger brain. We built something different, and we built it ourselves. Salient 1.0 is our own AI model. Our team trained it in-house, then locked it in place, so it never quietly changes and you can always explain what it did.
It runs on one high-end graphics card, not a data centre full of them. It already beats leading AI models on India’s rules and regulations, performs strongly on everything else, and offers one thing no other AI offers at any price: proof behind every answer.
Brain. Frozen 27B, multimodal, one GPU.
Beats leading AI models on India
Scores 20.8 points higher than a leading AI model on Indian tax and regulatory questions. That’s not a bigger model — it’s knowledge we built ourselves.
Proof, not guesswork
Every quote is checked word-for-word against its source before it’s shown to you. We tried six times to sneak a fake quote past this check. None got through.
A score, not a guess
The AI gives a confidence score with every answer, in a fraction of a second. If it isn’t confident, it says so or asks for a person — it never just guesses.
Ours, not rented
We trained it ourselves and we own it outright — the model, the training process and the roadmap. No dependence on a foreign company’s pricing or access.
Runs on one graphics card
The entire system runs on a single high-end graphics card — the kind built for serious AI work — in your own office. No need to build a data centre or hire a specialist infrastructure team.
Independently tested
Every number on this page comes from tests we built and run ourselves — not someone else’s leaderboard. You’re welcome to test it yourself, too.
India is where we proved it, not the limit of what it can do. The same system can be retrained for any other country’s laws, or for your company’s own area of work, in weeks rather than years.
The reasoning engine
Salient 1.0The reasoning engine that hands you the receipts.
Ask Salient about Indian tax law and it doesn’t recall. It retrieves, verifies character by character, and hands you the source. Ask it to write compliance code and it doesn’t guess. It runs the code and shows you the tests passing. Ask it something it doesn’t know and it says so, in seconds, with a confidence score attached.
A bigger model, or a fine-tune that quietly forgets what it knew and cannot explain what it learned.
Our own model, post-trained in-house on an open-weight base, then frozen and wrapped in twelve engineered layers that retrieve, verify, execute and decide, with a receipt on every claim. Hostable on a single consumer GPU.
Knowledge they don’t have
Indian statutory material is barely in the frontier models' training data. We built that corpus: 278,000+ verified statutory passages, 62,000+ regulatory pages refreshed every single day from RBI, SEBI, TRAI and beyond, and 3,556 pieces of compliance code that have been executed and tested.
Verification, not trust
Quotes matched byte-for-byte against the original before release. Anything with real-world consequence passes a server-side gate that text cannot argue with. Every answer ships a tamper-evident receipt naming the corpus version, the model build, and the exact bytes behind each claim.
Sovereign economics
One graphics card. No per-token bill, no foreign cloud, no third-party inference API in the loop, no outbound calls at runtime. India-resident by construction; DPDP-aligned by architecture — the same construction holds under GDPR, PDPA or any other jurisdiction’s data-residency rules.
The twelve layers
Under the hoodA frozen brain in an engineered suit.
Verified knowledge ships as signed, versioned packs. Swap the pack and you swap the domain or the country: legal today, healthcare tomorrow, India today, another jurisdiction next, same engine. Nothing is retrained, so nothing is forgotten and everything stays explainable.
And because the power lives in the suit rather than the wearer, a better model is a configuration change, not a rebuild. When the next open-weight model arrives, Salient re-fits to it overnight, and every pack, every receipt and every verified line of code carries straight over.
- 01BrainFrozen 27B, multimodal, one GPU
- 02LibrarySigned, versioned knowledge packsCorpus
- 03LibrarianHybrid search + reranking, 2.95× faster
- 04CoachEffort that scales with difficulty
- 05Fact-CheckerByte-for-byte quote verification
- 06Bouncer & ReceiptServer-side gate, tamper-evident provenance
- 07MemorySearchable past 64K, works at 5M
- 08Freshness RingDaily ingest, isolated regulated laneCorpus
- 09Coder’s Toolbox3,556 tested routines + sandbox
- 10SwitchboardRoutes general, verified and code
- 11Control RoomFull provenance, self-tuning optimisers
- 12Decision ModeCalibrated confidence in milliseconds
Layers 1, 3–7 and 9–12 are the engine: model, retrieval, verification, execution, memory, decision-making. They don’t change by geography. Layers 2 (Library) and 8 (Freshness Ring) are the corpus: India’s statutory and regulatory material ships today. A new country’s statutes, or a customer’s own proprietary domain, is the same ingestion pipeline pointed at a new source: weeks, not years, scoped as its own engagement.
The model and the memory
Salient 1.0180,000 tokens live on one graphics card. Searchable all the way to 5M.
Our own model: a 27B open-weight base, post-trained in-house, then frozen. Never retrained again once deployed, so nothing is forgotten and everything is explainable. Reads images and documents. 180,000 tokens of live context on a single consumer graphics card, no data-centre cluster required. No OpenAI, Anthropic or Google in the loop; no outbound calls at runtime; no per-token cost structure.
Verified knowledge as signed, versioned packs. Hybrid search with precision reranking and evidence in document order, re-engineered to 2.95× faster while getting more accurate, not less. Past 64,000 tokens a conversation becomes something Salient searches rather than strains to remember, and it keeps working at 300K, 1M and 5M tokens, where conventional context windows simply stop functioning.
Searched, not strained
Beyond the live window, Salient searches what it has read (hybrid search, precision reranking, evidence kept in document order) instead of straining to remember it. It keeps working at 300K, 1M and 5M tokens, where conventional context windows stop functioning.
278,000+
verified statutory passages
62,000+
regulatory pages refreshed every day from RBI, SEBI, TRAI and beyond
3,556
pieces of compliance code, executed and tested
100B
tokens in the indexed library
Decision Mode: the newest power
Layer 12A number, not a vibe.
- A confidence score on every answer — a number, not a vibe
- Low-confidence answers escalated or flagged, never shipped
- Instant “I don’t know” — seconds, not a wasted minute
- Agents checked at every step without slowing them down
- Read from probabilities, not written out — milliseconds, at a fraction of the compute
Every claim on this page is scored on our own Harness: evaluation engineering we built in-house, not a borrowed leaderboard. You’re welcome to rerun it against your own tests.
How to evaluate it
- Pick one costly, repetitive workflow — not a whole department.
- Point Salient at your historical cases and outcomes, and run it in parallel with your team on our Harness.
- Measure the delta in accuracy, turnaround time and cost against your baseline. You keep the results either way.
Deployment and data posture
- On-premises, air-gapped if required
- Your own cloud tenancy, under your keys
- Hosted by us, in the country your rules require
- Regulated data stays inside your boundary at all times
- No client data trains anything outside your instance
- Complete network isolation on the regulated lane — proven, not promised
- Full audit logging; HIPAA controls, BAA on request
Umpire’s AI firewall — included on every Salient deployment
About Umpire- Content filtering
- Denial of topics
- Keyword filters
- Prompt injection prevention
- Malicious link protection
- Image content filtering
- Language detection
- Gibberish detection
- Toxicity detection
- Bias content detection
- Sentiment detection
- PII & sensitive-data detection
- DLP
- Document content scanning
- OCR
- Code detection
- Real-time detection
- Regex-based detection
- LLM-based detection
- Contextual grounding
- Answer relevance
- Token limit
- Input rate limiter
- Input- and output-level protection
- Block & audit mode
- SIEM integration
- User activity logs
- Compliance mapping — ISO 42001, MITRE ATLAS, NIST
- Protection from AI chatbots over browsers
Built for your industry — whichever it is
- Banking & financial institutions
- Insurance
- Capital markets
- Healthcare & life sciences
- Government & public sector
- Manufacturing
- Aerospace & defense
- Logistics & supply chain
- Retail & e-commerce
- Telecom
- Energy & utilities
- Real estate
- Tax & legal
Salient is frontier-class wherever India’s documents decide the answer today — proof the same engine, retrained, reaches frontier-class wherever your documents decide the answer next. We aren’t betting on one model. We’re building the thing every model will need.