KRIYA CORTEX

Developed by Dr. Akhilesh Vats at Acme Research Solutions

See how it thinks

Evidence before prose. A reference no tool returned is never written.

Checks without AI. Quotes, citations, arithmetic and word limits verified by code.

Local-first. Chats, memory and files stay on your machine.

Palm-leaf manuscript folios drawn in code, representing verified references

Ledger grid drawn in code, representing deterministic checks of quotes, citations and arithmetic

Kolam dot pattern drawn in code, representing local-first memory

The research engine that checks its own work.

How it thinks

The request is rewritten into an explicit specification: audience, length, required and forbidden phrases, and whether outside evidence is needed. Closed-book questions and pure text transformations bypass retrieval.

Five chakras in a row, representing the five-stage research pipeline

01 · Interpret

02 · Retrieve

Web search and six scholarly indices run concurrently: PubMed, Europe PMC, arXiv, Crossref, Semantic Scholar and OpenAlex. Records without abstracts and retracted papers are discarded; the rest are re-ranked by dense-embedding similarity.

A writer stage composes against an evidence packet with fixed source identifiers. A separate reviewer then audits citations, quotes, arithmetic and the stated requirements, and edits the draft itself, logging every change as add, remove or correct.

With no model in the loop, each quote is matched to its source sentence, every citation ID is resolved, including grouped forms such as [3, 5] and [3–5], and word limits and required phrases are enforced. If QA fails, the draft is still delivered and marked for review.

03 · Draft and independent QA

04 · Deterministic verification

Features

Nine mechanisms, each described from the code that runs it.

#1 · Reference integrity

Every citation climbs a verification ladder: DOI to the canonical Crossref record, PMID to Europe PMC and PubMed, title match at 0.60 similarity or above. A reference no tool returned is written as [complete: reference], never recalled from memory.

Palm-leaf manuscript folios drawn in code, representing reference integrity

#2 · Forced computation

ANOVA with Tukey HSD, Dunnett, 4PL IC50 with 95% CI, release kinetics ranked by AIC, f2 similarity, ICH Q1E shelf life and Arrhenius run from a tested Python library. When a message carries data, the first step must execute code.

Word and PDF files are edited in place as atomic transactions: if one operation fails, nothing is written. Replacements are verified by searching the saved file, fonts are matched, and the original is never overwritten.

#3 · Surgical Word and PDF editing

Surya disc with a fitted curve drawn in code, representing pharma statistics computed in Python

Jaali lattice drawn in code, representing in-place Word and PDF editing

#4 · Audit with a correction ledger

A deterministic deep scan flags count–percent mismatches, impossible p-values and conflicting IC50, PDI or zeta values. Every number is then recomputed, every reference verified, and corrections are applied only from a saved ledger.

Ledger grid drawn in code, representing document audits with a correction ledger

#5 · Evidence-gated memory

A fact is stored only if its supporting quote and every number in it appear in your own words. Memories carry provenance labels and are retrieved by fusing keyword and semantic search. Workplace folders keep a running brief.

Questions are planned as four to seven research angles, gathered in waves, gap-checked and merged into one renumbered source list. Ten local OCR engines, including Hindi, read scanned pages.

#6 · Dig: multi-angle research

Kolam dot pattern drawn in code, representing evidence-gated memory

Stepwell drawn in code, representing multi-angle research

#7 · Desktop agents with approvals

Excel, Word, Windows controls and an isolated browser. Plans are approved first, a baseline is captured before every write, and each side effect is journaled so an interrupted step is never replayed.

Twenty-four-spoke chakra drawn in code, representing desktop agents with approvals

#8 · Cortex Code

An architect stage plans modules with acyclic dependencies; files are built in parallel and parser-checked; a repair agent fixes failing tests without touching them; a headless browser runs a visual review.

Paise-exact arithmetic, CGST + SGST or IGST by place of supply, gap-free series within CGST Rule 46(b), GSTIN checksum and HSN/SAC validation, with a print-ready PDF and an editable Word file.

#9 · GST invoices and quotations

Nagara temple shikhara drawn in code, representing Cortex Code builds

Bahi-khata ledger drawn in code, representing GST invoices and quotations

Built not to repeat a mistake

Every tool action carries a fingerprint of its input. A write whose outcome is uncertain is never repeated automatically.

Jantar Mantar instrument drawn in code, representing reliability by design

One reply at a time

Completion gate

A requested Word or Excel file counts as delivered only after it is saved, read back and verified.

Web pages, documents, OCR text, library excerpts and memory reach the models as data, never as instructions.

Local servers answer only the same machine and page; automation also needs a per-session token.

Untrusted content stays data

Loopback only

Request early access

A hosted version of KRIYA Cortex, with sign-in for phone and tablet, is on the way. Join the list for early access.

Contact

KRIYA Cortex is built by Dr. Akhilesh Vats at ACME Research Solutions, a GMP and ISO 9001:2015 certified pharmaceutical CRO in Baghpat, Uttar Pradesh.

Email

info@kriyacortex.com

support@kriyacortex.com