Muhammad Atashi
Evidence-grounded analysis · multi-format ingestionPRIVATE COREV2.1.2

Universal AI Analyzer

A domain-neutral Python service that snapshots heterogeneous sources, extracts and profiles them deterministically, runs hierarchical source-grounded analysis, validates every cited source/chunk relationship, and renders bounded Telegram reports.

Owns
Grounds multi-format extraction and hierarchical synthesis
Status
Source-grounded analysis system · bounded multi-format evidence pipeline
Stack
Python · Pydantic · LLMs · AsyncIO · Telegram · Google APIs

Source is not publicly released; claims are bounded to the documented release evidence.

Hard problem

If the source cannot be read, the system should refuse to invent an answer.

Question · Can this conclusion be traced to evidence the system actually read?

APPLIES TO

Analysts · Operations teams · AI Engineering teams · Multi-format evidence workflows

PROOF

Multi-format source-grounded analysis · 64 targeted tests · 500-report renderer fuzz · 48/48 wheel RECORD verification

CAPABILITIES

10 declared in the project record

  • Multi-format source ingestion
  • Bounded snapshots + SHA-256 provenance
  • Deterministic table profiling
  • Stable source/chunk identity
  • Hierarchical analysis
  • Strict schema validation
  • Citation ownership checks
  • Bounded semantic repair
  • Adaptive report rendering
  • Fail-closed unreadable-source policy

Source provenance

Synthesis never gets to invent the evidence graph.

  1. 01

    Sources

    text · tables · documents · structured
  2. 02

    Ingest

    API + file surfaces · bounded snapshots · SHA-256
  3. 03

    Extract

    format-aware parsing · deterministic profiling
  4. 04

    Chunk

    stable source/chunk identity
  5. 05

    Analyze

    hierarchical grounded synthesis
  6. 06

    Validate

    schema · citation · ownership · coverage
  7. 07

    Report

    controlled bounded output
  8. 08

    Deliver

    adaptive Telegram page splitting

FAIL-CLOSED SOURCE

If the source cannot be read, it cannot become evidence.

Fixtures
25
Accepted
24
Rejected
1

OCR/Vision for scanned PDFs/images is not supported.

500 reports · 3,816 pages · max 3,638 UTF-16 units

Engineering decisions

The constraints that matter.

01

Source identity stays outside model authority

The model may synthesize evidence but cannot invent the evidence graph it is allowed to cite.

02

Compute deterministic facts before synthesis

Structure, missingness, distributions, and provenance should not depend on generation quality.

03

Unreadable evidence fails closed

A source that cannot be extracted should never become plausible-looking analysis.

Measurement contract

What would make the preferred answer fail?

EVALUATION METHOD

Analysis quality starts with source fidelity: unreadable, uncited, or ownership-invalid evidence must fail before synthesis is accepted.

Format matrices, malformed-input suites, source/chunk mutation tests, deterministic profiling fixtures, renderer fuzzing, and provider-failure replay.

01

Format coverage

accepted + expected rejection matrix
02

Extraction fidelity

content preserved through parsing
03

Citation validity

source/chunk ownership checks
04

Profile accuracy

deterministic table statistics
05

Renderer safety

bounded output under fuzz
06

Failure recovery

provider/schema/repair behavior

Validation

What the evidence actually supports.

Claims stay scoped

Targeted regression/security/fidelity tests

VERIFIED

64 passed

Dependency-available release checks; not represented as the full pytest/coverage suite.

Source matrix

VERIFIED

25 fixtures · 24 accepted · 1 expected scanned-PDF rejection

OCR/Vision for scanned PDFs/images is not supported.

Compact offline E2E

SYNTHETIC

24 sources · 34 chunks · 5 simulated LLM calls · 2 Telegram pages

Simulated provider path; not live semantic-quality proof.

Long offline E2E

SYNTHETIC

24 sources · 34 chunks · 5 simulated LLM calls · 26 Telegram pages

Offline evidence only.

Renderer fuzz

VERIFIED

500 reports · 3,816 pages · max 3,638 UTF-16 units

Renderer boundary, not end-to-end delivery proof.

Wheel structure

VERIFIED

48 RECORD entries verified

Package structure verified; production deployment not implied.

Boundaries

— Not OCR/Vision by implication— Not a generic document chatbot— No uncited conclusions— No silent source/chunk ownership repair
← Selected systemsNext · Centris Property Monitoring