Enterprise FAQ: Buyer Readiness

Anweshna was founded by Kishore Bikomiya, who recognized that generic AI wrappers were failing institutional deal teams due to hallucinations and boilerplate blindness. He built Anweshna's deterministic pipeline specifically to meet the rigorous materiality standards of corporate development and private equity.

We have a dedicated team of 12 full-time employees, consisting primarily of AI researchers, senior software engineers, and former diligence analysts.

Our core engineering, research, and security team is headquartered in London, UK.

We are proudly bootstrapped and revenue-funded. This financial independence ensures that our product roadmap answers exclusively to our clients' strict security and accuracy needs, rather than the rapid growth mandates of venture capital investors.

We serve a growing, exclusive roster of mid-market private equity firms, corporate development teams, and boutique M&A advisories globally. Exact numbers and references are available under NDA during the Enterprise procurement process.

Our platform has successfully processed hundreds of live M&A transactions, parsing and analyzing millions of pages of confidential deal documents.

Our benchmarks show that 78.7% of our AI findings are independently verified by our secondary deterministic verification layer. The remaining findings are explicitly flagged as "Unverified" for human review.

In our latest 11-filing corporate benchmark run, our deal-killer false-negative rate was 0.0% (14 out of 14 critical issues surfaced successfully). We prioritize recall at the blocking gate to ensure material risks are never missed.

Our false-positive rate on the primary blocking gate is 15.4%. This demonstrates our intentional design choice to bias towards over-reporting rather than missing a material risk.

Our unverified rate—which represents the absolute upper bound on hallucinations—is strictly measured at 11.5%. Findings without an exact verbatim match in the source document are immediately flagged to the user.

Our infrastructure routinely scales to process Enterprise data rooms containing over 50,000 pages of highly unstructured, multi-format documents.

Our proprietary cross-document reconciliation engine (cross_doc) automatically cross-references entities, dates, and definitions across the entire data room to flag contradictory clauses (e.g., a term sheet definition conflicting with the final SPA).

We utilize enterprise-grade PyMuPDF extraction with fuzzy matching and error tolerance. When text is completely garbled or illegible, the system gracefully degrades by flagging the specific page for manual human review rather than attempting to guess the contents.

We currently do not explicitly support cursive handwriting analysis. Heavily degraded scans that fail OCR confidence thresholds are flagged as unreadable and skipped to prevent hallucinated analysis.

Our SOC 2 Type I readiness assessment is currently in progress, with final audit completion targeted for Q1 2027.

We are actively laying the groundwork for our Information Security Management System (ISO 27001) and AI Management System (ISO 42001), targeting formal certification within 12 to 18 months.

Default processing and storage occur in the US (AWS us-east-1). Strict EU/UK data isolation and regional processing are available on our Enterprise plans.

We maintain a strictly limited subprocessor list: Anthropic (LLM Inference), Supabase (Database & Auth), AWS (Infrastructure), and Paddle (Payments). See our Security Page for details.

Documents deleted by a user (or via retention policies) are immediately hard-deleted from active object storage and the database.

Yes. Deleted data ages out of our rolling, encrypted database backups within a maximum of 30 days, ensuring complete eradication.

We commit to notifying affected customers within 48 hours of discovering a confirmed security breach, as enshrined in our standard Data Processing Agreement.

Yes. Custom audit rights, security questionnaires, and architecture reviews are fully supported for customers on our Enterprise tier.

We currently utilize the Claude 3.5 Sonnet model family via the enterprise API, chosen for its optimal balance of deep reasoning, large context windows, and low latency.

Absolutely not. We process documents under Anthropic's enterprise terms that contractually prohibit using any client inputs or outputs to train or improve their foundation models.

All extracted document text passes through a PII scrubber and is strictly treated as untrusted user data. We enforce rigorous system prompts and XML-bounded injection constraints before the text ever reaches the LLM.

Every extracted finding is mathematically linked to its source via internal document_id and bounding-box page tracking, allowing 1-click verification of the original text directly in the platform.

We employ a second, independent model (the evidence_verifier) that attempts to prove the first model's claims against the raw, extracted text. If it cannot definitively prove the claim, the finding is explicitly flagged as unverified.

Yes. All findings, including their corresponding exact-text citations and metadata, can be seamlessly exported to JSON, PDF, and DOCX formats for integration into external legal reports.