The lending
operating system.
Litehouse is an AI-native, open-source, on-premise operating system for India's small and mid-tier housing finance companies and non-bank lenders. It runs three production agents — Aryabhatta-Analytics, Aryabhatta-CAM, and Aryabhatta-Legal — on the lender's own infrastructure, under the lender's own board-approved AI policy, with an immutable audit lineage that survives an NHB IT inspection by construction.
Litehouse is not a SaaS. It is a software stack you install, a corpus you own, and a runtime you operate.
Why an operating system,
not software.
The horizontal AI vendors are building applications. Applications sit on top of someone else's data, someone else's model, someone else's compliance posture. For the chairman of a ₹400 Cr HFC who watched IL&FS unwind in September 2018 and DHFL in 2019, the question is not "will this AI work?" — it is
"who controls it when the regulator walks in?"
Litehouse answers that question architecturally. The model weights live on your hardware. The borrower data never leaves your perimeter. The audit log is yours. The board policy is yours. The Consent Manager integration is yours. We ship the operating system; you operate it.
Three production agents.
Each agent replaces a specific desk inside a small or mid HFC — and each ships with the corpus, rule engines, and audit hooks the desk's regulator will ask about.
Analytics
ARYABHATTA-ANALYTICS
Replace the SQL desk. Keep the analyst.
The SQL-bound analyst desk that takes 4 days to answer a portfolio question
Head of Credit, CFO
CAM
ARYABHATTA-CAM
The Credit Assessment Memo, drafted in minutes.
The 6–9-day manual CAM cycle, especially for informal-income borrowers
Head of Credit, Credit Committee
Legal
ARYABHATTA-LEGAL
Title vetting. SARFAESI. Section 138.
Outsourced title vetting, SARFAESI Sec 13(2) drafting, S.138 notices
CS, Head of Legal, Head of Recovery
Compliance, by construction.
A horizontal cloud-LLM stack achieves FREE-AI compliance by adding controls retroactively. Litehouse achieves compliance because its architecture is the controls. On-prem + open-weight + immutable audit + HITL + board-policy-on-install — the regulator's "show me your AI inventory and your model risk management" question is answered on the day of inspection, not 14 weeks before it.
All seven sutras, all six pillars. Board-approved AI policy template ships with the deploy package.
On-prem deployment trivially satisfies cross-border transfer restrictions; Consent Manager integration scaffolded.
AI inventory, model registry, change log, access trail — the four documents the inspector asks for — emitted by the runtime.
Section 13(2) and 13(4) drafting templates, vetted by the Aryabhatta-Legal corpus.
Cheque-bounce notice generation aligned to Sanjabij Tari v. Kishore S. Borcar (2025) procedural directions and Damodar S. Prabhu compounding guidelines.
Borrower-wise NPA + ECL Stage 1/2/3 with dual-auth governance — MCV is the structural primitive.
Forty-eight profiles.
In an affordable HFC underwriting a tailor in Virar with no GST, no salary slip, and a one-room kirana downstairs — the bottleneck is not the math. It is the surrogate-income reasoning. Aryabhatta-CAM ships with the recipe for every informal-income household in semi-urban India.
Five commitments.
MCP-native. Hybrid rule + LLM. Human-in-the-loop. Air-gappable. Open-weight. Five non-negotiables that decide what we ship and what we refuse to.
- 01MCP-nativeBuilt on Model Context Protocol (Linux Foundation AAIF). Vendor-neutral, governed, open-standard agent fabric. Every connector is portable.
- 02Hybrid rule + LLMDeterminism first, intelligence second. Numeric aggregations, regulatory checks, date arithmetic are deterministic. LLM for reasoning, narrative, translation.
- 03Human-in-the-loop, by defaultEvery action with regulatory weight routes to a human approver. Configured per board-approved AI policy. The agent owns the draft; the human owns the decision.
- 04Air-gappableThree deployment modes — fully on-prem, sovereign Indian cloud, public cloud in private VPC. No telemetry leaves perimeter without approval.
- 05Open-weightModel registry ships with audited, production-grade open-weight models — Qwen 3, Llama 3.3, Qwen 2.5 Coder, Mistral Small. Lender pins specific weight hash. No closed-API dependency.
Open-source. Strong on Devanagari, Tamil, Kannada, Bengali, Gujarati, Marathi, Punjabi, Malayalam.
Built-in PAN, Aadhaar recognisers; custom IFSC, GSTIN. PII redaction & pseudonymisation by default.
SARFAESI · NI Act · RERA · ToP · state stamp acts · SC/HC precedents (Mardia, Transcore, Celir LLP, Sanjabij Tari).
The code is free.
The expertise is paid.
Apache 2.0 (model registry components governed by underlying open-weight model licenses). Repo at github.com/aryabhattalabs/litehouse. The MongoDB / HashiCorp / Elastic / GitLab / Confluent playbook, adapted for Indian lending.
Full Litehouse runtime · all 3 agents · MCP fabric · deployment scripts · Discord/Slack support.
Community + named SLAs (Sev-1 < 4hr) · private support channel · monthly office hours · security advisory feed.
Supported + remote operations · model upgrade management · on-call DevOps · FREE-AI evidence pack management.
Managed + dedicated solutions engineer · custom-corpus extensions · NHB-inspection accompaniment.
PRICING ILLUSTRATIVE · LITEHOUSE PRE-SEED PROJECTION
The four-option
table.
For a ₹500 Cr AUM HFC, over three years.
Litehouse Managed is 78% cheaper than the horizontal vendor stack over three years for the same AUM bucket — and architecturally delivers what the horizontal stack cannot: sovereignty, FREE-AI compliance, and the 48-profile corpus.
| OPTION | 3-YR TCO | WHAT YOU GET | WHAT YOU DON'T |
|---|---|---|---|
| Status quo | ₹3.6 Cr | Manual analyst desk + outsourced legal + cloud-vendor BI. Familiar workflow. | No AI leverage; FREE-AI gap; analyst attrition; Sec 138 windows missed. |
| Horizontal AI vendor | ~₹2.4 Cr | Chat layer on top of existing data; cloud-LLM dependency. | Cross-border data exposure (DPDP friction); no Indian credit ontology; no 48-profile corpus. |
| Litehouse Managed ★ | ~₹1.5 Cr | Three production agents · on-prem · FREE-AI by construction · remote ops included. | Lender hosts hardware (one-time RTX 6000 Ada: ~₹40L). |
| Litehouse Enterprise | ~₹3.6 Cr | Managed + custom corpus + regulator liaison + NHB-inspection accompaniment. | Highest sticker; reserved for >₹2,000 Cr AUM. |
The Agrim pilot.
Reference, not revenue. Proof of architecture, not proof of scale.
Agrim Housing Finance Pvt Ltd. Incorporated December 2018, registered with NHB. AUM growth documented by CARE Ratings: ₹117 Cr (Mar 2023) → ₹285 Cr (Mar 2024) → ₹314 Cr (Jun 2024). Co-lending partners: Godrej Housing, Singularity Creditworld, Vastu, DMI, Gosree. Fintech Startup of the Year, IFTA 2023. Home/Mortgage Lender of the Year, India Credit Risk Management Summit 2024.
Architecture lock; RTX 6000 Ada delivered; Litehouse v0.9 installed; board-policy template adopted.
Aryabhatta-Analytics live. Replaces 80% of analyst SQL queries. Baseline TAT measured.
Aryabhatta-CAM (10 of 48 profiles in production). Underwriter time saved measured.
Aryabhatta-Legal: title vetting + Section 13(2) drafting in production. NHB-inspection-ready audit log validated.
FREE-AI evidence pack. Board-policy + AI inventory + DPIA + algorithmic-audit kit.
A guide for the head of IT.
A single RTX 6000 Ada workstation amortised over 4 years costs the lender less than 6 months of a horizontal AI vendor's SaaS subscription — and the lender owns the asset, the model weights, and the borrower data lineage at the end.
| TIER | HARDWARE | CAPEX | OPEX / YR | FITS |
|---|---|---|---|---|
| Workstation | 1× RTX 6000 Ada (48GB) · 256GB DDR5 · 8TB NVMe | ~₹8L | ~₹1L | Small HFC · ₹100–500 Cr AUM |
| Single-node server | 2× H100 (80GB) or 4× L40S · 512GB · 16TB NVMe | ~₹40L | ~₹4L | Mid HFC · ₹500–3,000 Cr AUM |
| Cluster | 8× H100 + secondary node | ~₹2 Cr | ~₹15L | Large HFC / NBFC; group-shared infra |
| Sovereign cloud | Yotta Shakti / E2E reserved | Opex-only | ~₹80L–₹3 Cr | Lenders without DC capacity |
| Public cloud (private VPC) | AWS / Azure / GCP India region | Opex-only | ~₹1.2–4 Cr | Existing public-cloud commitments + DPO sign-off |
Run your HFC
as code.
Schedule an architecture briefing. We'll walk your head of IT and head of credit through a topology that fits your AUM and your inspection calendar.