You can deploy this model and start earning money today!

pookieca/
pookieai-qwen-3.6-27b

Name: pookieai-qwen-3.6-27b
Availability: InStock

Views

No views yet

Pookie AI

The conversational tutor for India's professional examinations.

Every year, more than five million Indians sit for a professional exam — Chartered Accountancy, Company Secretary, Cost & Management Accountancy, Bar examinations, judicial services, banking officer exams, securities-market certifications, and more. They study from overlapping syllabi, pay overlapping coaching fees, and ask the same questions in slightly different shapes.

What this repo is

pookieai-qwen-3.6-27b is the cloud-served conversational layer of Pookie CA — the first product in the Pookie AI line. Weights are not uploaded yet; this card stakes the namespace, locks the architecture, and marks the start of the public roadmap. v1 ships the first weights with full evaluation.

Sibling cards for the on-device tier:

pookieai-gemma-4-e4b-it-mtp-lora-unsloth — field-validated on iPhone 14 + A15 Bionic with 6 GB RAM. Android on-device deployment is parked pending flagship-tier (Tensor G4 / SD 8 Gen 3, ≥12 GB) validation — see that card for the device-by-device field-testing matrix.
pookieai-gemma-3-1b-android (Abandoned:- unreliable at prod) — proposed floor model for sub-6 GB Android once field-validated. Not yet tested on real devices.

The thesis

CA Final's syllabus is the closest thing India has to a superset of professional finance, law, and commerce knowledge. Adjacent professional certifications are slices of it with bounded pattern shifts.

Validated overlap percentages (against existing CA Final + SPOM Law corpus):

Adjacent exam	Subject area	Overlap
IBC Insolvency Professional / LIE / Registered Valuer	IBC, Companies Act, SEBI valuation	85–90%
CS Executive (ICSI)	Companies Act, SEBI, FEMA, IBC, Economic Laws	70–90% per paper
CMA Intermediate (ICMAI)	Cost, Direct Tax, GST, FR, Company Law, Audit	60–85% per paper
RBI Grade B Phase 2 Finance & Management	Financial markets, derivatives, governance, accounting	55–65%
NISM Series V-A / X-A / VIII	MF taxation, portfolio theory, equity derivatives	35–60% per series
ACCA Applied Skills (FM / PM / AA)	Financial management, performance management, audit	50–80% per paper

A unified authored content layer, indexed by canonical reference (Section / Standard / Rule / Regulation), serves every one of these surfaces with per-exam adapters on top. That's the architecture.

How Pookie AI is built

A [pgvector+neo4j] spine plus a conversational narrator. Each layer does what it's good at.

The narrator (model)

Three behaviours, trained for:

Answers cite their source. Every response is grounded in a retrieved passage from the authored corpus. When retrieval is empty, the model says so — it does not improvise.
Numbers stay exact. Marks, section numbers, due dates, amendment cut-offs are passed through verbatim from the spine, never paraphrased or estimated.
The voice is Indian. Indian numbering (lakhs, crores), ₹ for currency, and measured Hindi-English code-mix reserved for memory aids — never in headings, navigation, or instruction.

These three together are the entire job.

Model roadmap

Version	Stage	What lands
v0	This card	Namespace staked, architecture locked, public commitment
v0.1	Next	RAG-only baseline shipping inside Pookie CA. No fine-tune. Failure modes logged publicly
v0.2	Conditional	LoRA pilot if v0.1 misses the discipline targets. Eval card with refusal rate, citation accuracy, voice consistency
v1	First weights	Full eval, training-data lineage, reproducibility notes
v2	Lineup	Cloud 27B and on-device 1B sibling co-released for matched product experience across phones

No forward-looking dates. The changelog logs work as it ships.

Get in touch

Product: pookieca.com
Partnerships: coaching faculty, content authors, eval contributors, and institutional collaborators — open a Discussion on this repo.
Engineering deep-dives: the architecture documentation (deterministic spine, retrieval indexing, per-exam adapters) will be published alongside v1.

pookieca/pookieai-qwen-3.6-27b