Sovereign Industrial AI Workbench. Scattered plant knowledge becomes an authorized, traceable decision flow.
Self-hosted and air-gapped: it runs on the plant's own GPU server and nothing leaves the premises.
Website: yukti.cypherion.tech
Version 0.5.1 · Windows 10 / 11 · macOS 13.3+ on Apple silicon · English, हिंदी, ಕನ್ನಡ · Documentation
| ⚡ ~1 ms | 🚀 44–58 tok/s | 🌏 3 languages | 🔒 0 bytes out |
|---|---|---|---|
| Laya routes a request before any LLM call |
Gemma 3 4B on a 6 GB laptop GPU |
English · हिंदी · ಕನ್ನಡ with offline speech |
Offline guard blocks and counts egress |
📑 Contents · The problem · The answer · End to end · Features · Architecture · Why it holds up · Performance · Get started · Roles · Security · Documentation · Roadmap · Team
02:15 AM — pump A2 trips. The electrician needs ratings, failure history, open work orders, wiring, isolation steps and the approved troubleshooting procedure. Now.
flowchart LR
T(["⚠️ Pump A2 trips"]) --> E["👷 Electrician<br/>needs 6 answers"]
E -.-> A["Equipment master"]
E -.-> B["Maintenance history"]
E -.-> C["Electrical drawings · P&IDs"]
E -.-> D["Manuals · SOP rev 2 vs rev 3"]
E -.-> F["Scanned inspection reports"]
E -.-> G["Emails · shift logs"]
style T fill:#e5484d,color:#fff,stroke:#e5484d
Structured records, PDFs, old scans, conflicting revisions — and cloud AI is off-limits for P&IDs and financials. So people search manually, or paste confidential data into public tools.
Yukti is a local, role-aware operating layer that connects the plant's knowledge, people, workflows and decisions — without sending confidential information outside the organisation.
| 🧭 Model auto-selection | 🤖 Agentic, multimodal work | 🛡️ Sovereign & governed |
|---|---|---|
| Laya (ONNX, ~1 ms) returns typed decisions: intent, department, urgency, worker | Plans multi-step work with local tools: files, spreadsheets, search, sandboxed code | Runs on the plant's own GPU server, air-gapped |
| A durable queue hands each task to the right specialist | Scans, handwriting, drawings and photos go through OCR + VLM | An ABAC policy engine decides access — never the LLM |
| New open-weight models plug in without a redesign | Delivers approval notes, Word / PowerPoint / Excel, calculations with steps | SHA-256 hash-chained audit of every action |
flowchart LR
Q(["❓ A question<br/>or 📄 a scan"]) --> I["🪪 Identity & policy<br/>post · department · clearance<br/>asset scope · reporting chain"]
I --> L["🧭 Laya routes<br/>typed decision · ~1 ms"]
L --> W["📥 Durable queue<br/>specialist worker"]
W --> M["🧠 Tuned local LLM<br/>OCR + VLM · analytics · sandbox"]
M --> E["📚 Cited evidence<br/>known · missing · conflicting"]
E --> O(["✅ Answer · 📝 Note · 🔔 Alert"])
style Q fill:#0e5a7a,color:#fff
style O fill:#2ea44f,color:#fff
sequenceDiagram
participant U as 👷 Electrician
participant B as 🔥 Boiler engineer
participant P as 🛡️ Policy engine
participant H as 👔 Dept. manager
U->>P: A2 dossier?
P-->>U: ✅ allowed · cited dossier
B->>P: Finance audit?
P-->>B: ⛔ denied · outside role
B->>H: Access request (why)
H-->>B: ✅ time-bound grant ≤ 72 h
Note over P,H: decision recorded in the audit chain
| Field | Example |
|---|---|
| Source | SOP-EL-014 |
| Revision | rev 3 · CURRENT |
| Page / record | p.2 · A2_WorkOrders!34 |
| Timestamp | 2026-09-29 |
| Confidence | 0.84 |
| Status | 🟢 KNOWN · ⚪ MISSING · 🟠 CONFLICTING |
| Found by | bm25 tags graph vector |
Not merely an answer — an answer the right person is allowed to trust.
| Feature | In one line | |
|---|---|---|
| 🧭 | Laya + work queue | Small decision model picks the worker: policy · small LLM · OCR + VLM · analytics · 27B reasoning · coder · workflow |
| 📚 | Hybrid retrieval | BM25 + asset tags + metadata + knowledge graph + optional BGE-M3 vectors — all behind the policy check |
| 🧰 | Agentic workbench | Approval notes, Word, PowerPoint, Excel, calculations with every step, code in a no-network sandbox |
| 🧾 | Scan → finding | OCR + VLM → verify vs asset master → human review → finding → approval note → approver lane |
| ✅ | Approver lanes | Electrical · mechanical · HSE · procurement · finance · management, with acknowledgement, escalation, follow-up |
| 🏷️ | Asset ledger | Vendor, serial, location, owner, certificate, interval, warranty, expiry — alerts before anything lapses |
| 📈 | Production intelligence | Demand forecast + HiGHS optimizer → scenario card: proposed change, expected value, alternatives |
| ⚙️ | Performance Engine | CUDA/Vulkan auto-pick, VRAM auto-fit, q8_0 KV cache, hot KV slots (TTL + LRU), speculative decoding, validated profiles |
| 📷 | Ask with a photo | Reads tag plates and stencils offline, matches the asset register, points out rust and leaks (add-on) |
| 🎙️ | Speak instead of typing | Offline speech in English, Hindi, Kannada — whisper.cpp + AI4Bharat IndicWhisper (add-on) |
| 🧪 | Isolated research layer | Experimental models with their own guardrails — no documents, tools or approvals |
| 🌐 | Knowledge base | Contacts and approved points of contact, company history, digitised old records, MRPL public data (140 sources) |
 Cited answer with facts and policy decisions |
 Conflicting sources with a recommendation |
 Inbox: access requests and findings |
 Production intelligence |
 Knowledge: HOD uploads for own department |
 MRPL public intelligence with sources |
 Admin: users, roles, MFA |
 Admin: organisation-wide AI settings |
 Admin: measured usage and health |
 Desktop app setup |
flowchart TB
S(["📄 Scanned report"]) --> O["🔤 OCR<br/>text · tables · tags · dates"]
O --> V["👁️ Vision model<br/>stamps · symbols · markings"]
V --> C{"✔️ Verify against<br/>asset master"}
C -- "mismatch or<br/>low confidence" --> R["👤 Human review<br/>confirm · correct · dismiss"]
C -- "all fields verified" --> F["🧾 Finding<br/>asset · procedure · owner<br/>due date · evidence region"]
R --> F
F --> N["📝 Approval note<br/>(.docx)"]
N --> L["✅ Approver lane"]
L -- "decided" --> D(["Approved / rejected"])
L -- "not acknowledged in time" --> E["⏰ Escalated<br/>HOD → management"]
E --> D
style S fill:#0e5a7a,color:#fff
style D fill:#2ea44f,color:#fff
A forgotten scan becomes an owned workflow with a deadline.
flowchart LR
H["📊 Authorised history<br/>production · sales"] --> F["🔮 Forecast<br/>demand ± uncertainty"]
X["💹 Prices · crack spreads<br/>inventory · feedstock<br/>maintenance · storage<br/>dispatch · contracts"] --> O
F --> O["🧮 Constrained optimizer<br/>HiGHS LP"]
O --> C["🗂️ Scenario card<br/>% change · expected value<br/>assumptions · alternatives"]
C --> L["🧠 LLM explains"] --> P(["👔 Planners decide"])
style P fill:#2ea44f,color:#fff
🔒 Yukti never changes DCS or SCADA setpoints. The optimizer calculates, the LLM explains, planners decide.
flowchart TB
subgraph C["👥 Employee PCs"]
D["🖥️ Desktop app · Electron"]
BR["🌐 Browser"]
end
subgraph S["🏭 Yukti Server — plant GPU server · offline guard on"]
direction TB
subgraph L1["1 · Who is asking"]
A["argon2id + MFA · sessions · IAM attributes"]
end
subgraph L2["2 · Decide & route"]
P["🛡️ ABAC policy engine"] --> LY["🧭 Laya · ONNX"] --> Q["📥 Durable queue"]
end
subgraph L3["3 · Knowledge & governed output"]
R["📚 Hybrid retrieval"]
FA["⚖️ Facts · asset ledger"]
WB["🧰 Workbench · sandbox"]
SF["🧾 Scan → finding · lanes"]
PR["📈 Forecast + HiGHS"]
end
subgraph L4["4 · Performance Engine"]
E["🧠 llama.cpp CUDA / Vulkan · profiles · hot KV · watchdog"]
end
AU[("🔗 SHA-256 audit chain")]
end
D -->|plant LAN| A
BR --> A
A --> P
Q --> R & WB & SF & PR
R --> FA --> E
WB --> E
L2 -.-> AU
L3 -.-> AU
flowchart LR
A["1️⃣ Ingest & digitise<br/>PDF · DOCX/XLSX · scans<br/>text vs scanned · OCR + VLM<br/>link to asset tag"] --> B["2️⃣ Govern & retrieve<br/>IAM + policy engine<br/>hybrid retrieval<br/>revision-aware citations"]
B --> C["3️⃣ Route & execute<br/>Laya classifies<br/>queue dispatches<br/>checkpoint per chat ID"]
C --> D["4️⃣ Act, approve & audit<br/>notes · alerts · scenarios<br/>accountable human<br/>every step logged"]
| Layer | Stack |
|---|---|
| 🐍 Language & app | Python · FastAPI · React + Vite · Electron |
| 🧠 Local inference | llama.cpp · ONNX Runtime · AI4Bharat ASR · CUDA / Vulkan |
| 🤖 Models & solvers | Gemma 3 4B · RapidOCR · Laya · HiGHS LP · optional: 27B-class reasoning, BGE-M3 |
| 🔐 Data & security | SQLite FTS5 · ABAC policy · argon2id + MFA · SHA-256 audit |
Full design: docs/ARCHITECTURE.md · research behind it: docs/REFERENCES.md
| 🛡️ Policy, not prompts | 🧭 Small model decides first |
|---|---|
| Authorization comes from the host policy engine, so the LLM cannot grant itself permission. A genuine need goes to the department manager for time-bound, recorded access. | Guardrail denials, off-topic and need-to-know questions are answered without calling the LLM. The prompt is fitted to the context window: old turns and the least relevant documents go first; facts and the question always stay. |
| ☁️ Cloud AI assistants | 🛡️ Yukti workbench | |
|---|---|---|
| Data | must leave the plant | on-premise, air-gapped |
| Cost | pay per token | one-time hardware, tuned to the local GPU |
| Serving | generic, no tuning | q8_0 KV cache · speculative decoding · profiles |
| Trust | fluent answer, no source | cited: known · missing · conflicting |
Risks, rollout and viability → docs/FEASIBILITY.md
RTX 4050 laptop GPU, 6 GB:
| What | Measured |
|---|---|
| 🚀 Gemma 3 4B generation (llama.cpp CUDA) | 44–58 tok/s |
| 🧠 Gemma 3 4B + Whisper turbo together | 5.1 of 6 GB VRAM |
| 📷 Photo question (OCR + model) | 4–10 s |
| 🎙️ Speech to text, Whisper turbo on GPU | 0.3–0.4 s for a 4 s question |
| 🧭 Laya routing | ~1 ms |
| Backend · Gemma 2 2B Q8_0 | Generation | First token |
|---|---|---|
| llama.cpp CUDA 12.4 | 51.7 tok/s (3.9× CPU) | 0.2 s |
| llama.cpp Vulkan | 50.8 tok/s | — |
| CPU only | 13.4 tok/s | — |
No cluster, no cloud. Larger models (e.g. a quantized 27B-class reasoning model) use Performance Engine profiles that administrators validate on their own server; the dashboard shows only values measured there.
🪟 Windows 10 / 11
flowchart LR
A["⬇️ Yukti-Setup.exe"] --> B["🖱️ Double-click<br/>no admin rights"] --> C["📦 Downloads its AI once<br/>+ 📷 photos + 🎙️ voice<br/>3.9 GB · 5.3 GB with NVIDIA<br/>resumable · verified"] --> D(["🔑 Sign in<br/>works offline"])
- Windows SmartScreen? More info → Run anyway (installers are not code-signed yet).
- 📷 Photos and 🎙️ voice come with the first install; an admin can remove either from the home screen.
- No internet? Copy the offline package (
ops\release-site.ps1) by USB.
🍎 Mac (Apple silicon, macOS 13.3+)
flowchart LR
A["⬇️ Yukti-macOS-arm64.dmg"] --> B["📂 Drag to Applications<br/>first open: Privacy & Security<br/>→ Open Anyway"] --> C["📦 First start downloads<br/>AI model + 📷 photos + 🎙️ voice<br/>3.9 GB · resumable · verified"] --> D(["🔑 Sign in<br/>works offline"])
- Runs on the Mac's own graphics chip (Metal): llama.cpp for answers, whisper.cpp for voice, both inside the app.
- Not notarized yet: macOS asks once, System Settings → Privacy & Security → Open Anyway.
The self-contained Yukti Server package ships yukti-server.exe, llama.cpp (CUDA 12.4 + Vulkan), Gemma 3 4B with its
image module, whisper.cpp, a pre-trained Laya, the UI and the data — no Python, Ollama or vLLM needed (but usable if present).
| Goal | Run |
|---|---|
| This PC only | Yukti.exe (or server\Start Yukti Server.cmd) |
| Serve the plant LAN | server\Start Yukti Server (Plant LAN).cmd — allow TCP 8000 for the plant subnet only |
| Publish the download | ops\build-release.ps1 → ops\release-site.ps1 · Mac: ops/build-mac.sh → Yukti-macOS-arm64.dmg |
Optional models (MODELS.txt): 27B-class reasoning · BGE-M3 vectors · AI4Bharat IndicWhisper · research model.
Requires uv + Python 3.12 and Node 22.
cd backend; uv sync # backend
cd ..\web; npm ci; npm run build # UI (or: npm run dev on :5173)
cd ..; .\ops\start.ps1 # API + UI on http://127.0.0.1:8000
cd desktop; npm ci; npm start # desktop client (npm run dist → Yukti-Setup.exe)| Tests | Command | Needs |
|---|---|---|
| 🧪 Unit / API | cd backend; uv run pytest |
nothing (no GPU) |
| 🔁 Live end-to-end (86 checks) | uv run python tests/live/e2e_live.py |
running server + model |
| 🧰 Live workbench | uv run python tests/live/workbench_live.py |
running server + vision model |
| 💥 Stress / fault injection | uv run python tests/live/stress_live.py |
running server |
| 🌐 Web | cd web; npm run build |
Node 22 |
First start seeds demo employees, MRPL public data and the example documents (~30 s). Reset: delete data\store\.
→ Contributing · API
| Role | Can |
|---|---|
| 👷 Employee | Cited chat over what they may see · workbench · request access · inbox |
| 👔 Head of Department | + own department's documents · approve access · findings · scan review · ledger |
| ✅ Approver lanes | Decide their lane: electrical · mechanical · HSE · procurement · finance · management |
| 🏭 Plant manager | Approvals · escalations · audit · work queue · production |
| 🏛️ Top management | The only role with company-level information (finances, strategy) |
| 📈 Planner | Plant model · forecast · plan scenarios |
| 📦 Stores | Register purchased and installed items in the ledger |
| 🔍 Auditor | Audit log · chain verification · export |
| 🧪 Researcher | Isolated research layer only |
| ⚙️ Administrator | All LLM configuration · users · backups · policies — but no plant documents |
| 🚫 LLM never decides access | Host policy runs before retrieval and before every tool call — deny-overrides, default deny, fail closed |
| 🔌 Offline guard | Blocks and counts outbound connections: proof of no external calls |
| 🔗 Tamper-evident audit | Logins, denials, retrievals, tool calls, answers, approvals — hash-chained |
| 🎛️ Advisory only | No tool can write to DCS/SCADA; safety steps need accountable humans |
| 📦 Sandbox & research isolation | Code runs without network or system access; experimental models stay isolated |
→ Security Model · report a vulnerability: docs/SECURITY.md · bugs: juhilprogramming@gmail.com
| 👷 Using Yukti | 🛠️ Running Yukti | 🧠 Understanding Yukti | 📋 Project |
|---|---|---|---|
| User Guide | Deployment | Architecture | Changelog |
| Admin Guide | Security Model | Contributing | |
| API · API v2+ | Feasibility & impact | Security policy | |
| Research & references | Third-party notices |
flowchart LR
P1["🧪 Phase 1 · Pilot<br/>one unit's maintenance knowledge<br/>dossiers · SOPs · drawings<br/>identity-aware RAG + citations"] --> P2["🧾 Phase 2 · Workflows<br/>scan → finding · approval routing<br/>asset ledger · expiry alerts"]
P2 --> P3["📈 Phase 3 · Planning<br/>production intelligence<br/>with planners"]
P3 --> P4["🏭 Phase 4 · Scale<br/>plant-wide · template for<br/>other refineries & PSUs"]
- 🔏 Code-signed installers (SmartScreen warns today)
- 🔐 Certificate management — HTTPS works with
--tls-cert/--tls-key; issuing and renewal is left to the plant PKI - 🛡️ A refinery-specific guard model — studied only in the isolated research layer for now
- 🔗 Read-first connectors to DMS, CMMS and SAP
- 🐧 A Linux server package
- 📏 Evaluation on real plant documents and calibration of Laya on labelled refinery workflows
| ✅ Real (measured or sourced) | 🧪 Example (by Team UniMinds, badged) | ⛔ Not included (from the plant) |
|---|---|---|
| All software behaviour, settings, tok/s, audit hashes, policy decisions · MRPL public data from 140 sources | 18 plant documents (datasheet, SOP and drawing revisions, troubleshooting guide, E-310 and PSV-118 scans, 1998 letter, finance audit, MSDS, manual, shift log, work orders, asset register, contacts) and the assets they describe | Yields, prices, demand history and contracts (the planner enters them) · the real asset register · real documents · the HR directory |
Team UniMinds · Team ID 152431 · Smart India Hackathon 2026
SIH26117 — Sovereign On-Premise Agentic AI Workbench using Open-Weight Multimodal LLMs for Confidential Industrial Work
Theme: Smart Automation · Category: Software
| Juhil Modi (lead) | Saswata Das | Harshit Tiwari | Rohan Jangam | Ritesh Prajapati | Sanskriti |
|---|
Copyright © 2026 Team UniMinds. Proprietary, under the Yukti Proprietary Licence: the IP stays with Team UniMinds (per SIH rules), MRPL and the Ministry of Petroleum and Natural Gas get free lifetime use, the SIH organisers may evaluate it, and anyone else needs our written permission. See NOTICE and third-party notices.