NetAuditAI

A network configuration security auditor that refuses to guess.

Built for Smart India Hackathon 2026, problem statement SIH26155 (NTRO, Cybersecurity): AI-Driven Multi-Vendor Network Security Compliance Auditor.

NetAuditAI reads router, switch, firewall and cloud firewall configurations, answers 24 security checks on every one of them, and cites the exact configuration line behind every answer. When it cannot decide, it says so instead of passing the check. It fixes what it is sure about, proves every fix by rescanning a copy, and never runs a command on a device.

Upload or SSH-collect a config  →  cited PASS / FAIL / UNKNOWN per check  →  posture + coverage
                                →  verified fixes as device commands  →  PDF report, SARIF, signed audit ledger

---

Contents

• Why it is different
• Quickstart
• Demo
• How it works
• Reading the results
• Vendor support
• The 24 checks
• Measured accuracy
• More than a checklist
• Configuration
• Command line and CI
• What is stored
• Deploying safely
• Known limitations
• Repository layout
• Testing
• Documentation
• License

---

Why it is different

Most config auditors either support a handful of vendors with hand-written rules, or put an LLM in front of the file and trust what comes back. NetAuditAI does neither.

Principle	What it means in practice
Undecided is an answer	UNKNOWN and NOT_CONFIGURED exist so missing evidence is never silently a PASS. Posture is computed from decided checks only, and coverage says how much could be decided.
Every verdict is traceable	Framework requirement → check → vendor-neutral fact → configuration line → verdict. No verdict without a cited line (or a documented vendor default).
AI proposes, code verifies, a human confirms	The AI only looks at checks the engine left undecided, its citations are verified deterministically, and its answers are never scored until an administrator confirms them.
Learning without training	An administrator confirms how an unfamiliar line reads; that becomes a typed, validated recognizer that answers decisively on every later scan, with no AI call.
Fixes are proven, not generated	Fixes come from fixed recipes or from the same recognizer that read the line. Each one is applied to a copy and rescanned; it only counts if the check now passes and nothing else regressed.
Read-only toward devices	Live collection opens one SSH session with read-only commands. Nothing NetAuditAI produces is ever executed on a device.

---

Quickstart

Prerequisites: Python 3.10+, Node.js 18+. An AI key is optional; everything except AI explanations, chat and AI proposals works without one.

# 1. Backend  (http://localhost:8000, API docs at /docs)
cd backend
python -m venv venv
source venv/bin/activate            # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --reload --port 8000

# 2. Frontend, in a second terminal  (http://localhost:5173)
cd frontend
npm install
npm run dev

Open http://localhost:5173, choose New scan, and upload demo-sih/cisco_oneclick.cfg.

No UI needed? Scan from the command line:

cd backend
python -m app.cli scan ../demo-sih/ --fail-on high

Optional settings live in backend/.env (copy backend/.env.example) and frontend/.env. See Configuration.

---

Demo

Four demos in demo-sih/. Each goes from a bad first scan to posture 100 with 0 problems once the corrected file is rescanned. This is pinned by backend/tests/test_demo_fix_to_100.py.

File	First scan	How it gets to 100
cisco_oneclick.cfg	posture 41, 15 problems	One click. Deterministic Cisco recipes, no input needed.
paloalto_ai_human.cfg	posture 30, 10 problems	5 fixed by seed write-back (no parser), 5 by an AI-drafted or typed command, each verified on a copy.
unknown_vendor.cfg	no score, 3 suspected problems	Teach 3 lines in Adaptive learning, then 3 verified commands.
fortigate/ (3 files, one upload)	fleet posture 80, 13 problems	One click; the download is a .zip of 3 corrected files. Also shows a cross-device NTP mismatch and a closed attack path.

The two-minute judge path and the full walkthrough are in docs/demo.md. A browser test (cd frontend && npm run e2e) walks the same path.

---

How it works

flowchart TD
    subgraph S1["1 . Ingest"]
        UP["Upload a file"]
        SSH["Collect over SSH<br/>Netmiko / NAPALM, read-only"]
        MEM["Held in memory, max 2 MB<br/>never written to disk"]
        UP --> MEM
        SSH --> MEM
    end

    MEM --> DET{"Vendor detection<br/>+ grammar coverage"}

    subgraph S2["2 . Read: configuration to SecurityFacts"]
        PAR["Dedicated parser + vendor defaults<br/><i>decisive</i>"]
        TOK["Generic tokenizer<br/>JSON and Terraform flattened first"]
        REC["Recognizers from the store<br/>380 shipped + taught, <i>decisive</i>"]
        HEU["Lexicon heuristics<br/><i>provisional</i>"]
        TOK --> REC
        TOK --> HEU
    end

    DET -->|"Cisco IOS or FortiGate,<br/>grammar confirmed"| PAR
    DET -->|"any other vendor,<br/>or unverified"| TOK

    FACTS[("SecurityFacts<br/>predicate, value, scope,<br/>cited lines, assurance")]
    PAR --> FACTS
    REC --> FACTS
    HEU --> FACTS

    subgraph S3["3 . Decide"]
        CTL["24 checks on every configuration<br/>+ organisation policy"]
        SCORE["Posture + coverage<br/>decisive evidence only"]
        CTX["Attack paths, risk, drift,<br/>fleet checks, framework views"]
        CTL --> SCORE
        CTL --> CTX
    end
    FACTS --> CTL

    subgraph S4["4 . Learn: unknown vendors"]
        JUDGE["AI judge (optional)<br/>redacted, fenced, budgeted, cached"]
        VER["Deterministic citation verifier<br/><i>ai_verified: shown, never scored</i>"]
        HUM["Administrator confirms<br/>in Adaptive learning"]
        DB[("Recognizer store<br/>feeds step 2 on every later scan:<br/>decisive, no AI call")]
        JUDGE --> VER --> HUM
        HUM -->|"saves a recognizer"| DB
    end
    CTL -->|"UNKNOWN /<br/>NOT_CONFIGURED"| JUDGE
    CTL -->|"suspected FAIL"| HUM

    subgraph S5["5 . Fix: decisive FAILs only"]
        RCP["Confirmed vendor:<br/>deterministic recipe"]
        WB["Unconfirmed vendor:<br/>seed write-back or candidate command"]
        RESCAN["Apply to a copy and rescan:<br/>check passes, nothing regresses"]
        OK["Administrator confirms<br/>never executed on a device"]
        RCP --> RESCAN
        WB --> RESCAN
        RESCAN --> OK
    end
    CTL --> RCP
    CTL --> WB

    OUT["UI, PDF report, baseline JSON, SARIF<br/>+ signed, hash-chained audit ledger"]
    SCORE --> OUT
    CTX --> OUT
    OK --> OUT

Stage	What happens	Code
Ingest	Upload or SSH collection. Collection is restricted to private address space by default; credentials are used for one session and never stored.	app/collect/
Detect	The vendor is confirmed only when the file follows its grammar (coverage threshold 0.7). Look-alikes (NX-OS, ASA, Arista, FortiSwitch, mixed files) are reported as unverified and take the generic path.	app/parsers/detector.py, coverage.py
Read	Parsers, recognizers and heuristics all produce the same vendor-neutral SecurityFacts. This is the Security Baseline Model, downloadable from GET /api/scan/{id}/baseline.	app/parsers/, app/structure/, app/facts/
Decide	Every check runs on every configuration; the vendor only decides where facts come from.	app/controls/, app/analysis/
Learn	AI proposals and suspected lines become recognizers only through an administrator.	app/ai/judge.py, app/adaptive/, app/db/
Fix	Recipes (Cisco, FortiGate), seed write-back, vendor reference commands and candidate commands, each verified by rescanning a copy, then written as the device's own CLI in lockout-safe order.	app/remediation/
Report	PDF per device, executive summary, SARIF, Ed25519-signed hash-chained ledger.	app/reporting/, app/ledger.py

The full design, including the exact gates a recognizer must pass, is in docs/architecture.md.

---

Reading the results

Status per check

Status	Meaning
PASS	Evidence says the setting is secure, with a cited line or a documented vendor default.
FAIL	Evidence says the setting is insecure. One FAIL per failing scope (interface, VTY range, policy).
UNKNOWN	Something relevant exists but could not be decided (conflict, missing unit, unread syntax, AI proposal awaiting confirmation).
NOT_CONFIGURED	Nothing relevant was found. Never counted as a PASS.
N_A	Proven not to apply, for example no VPN on a device whose parser reads VPNs. Only a confirmed parser can prove this.

Assurance per verdict

Assurance	Source	Counts toward posture?
parser	Dedicated Cisco IOS / FortiGate parser	Yes
confirmed	Shipped seed recognizer or administrator-confirmed recognizer	Yes
default	Documented default of a confirmed vendor	Yes
heuristic	Lexicon reading of an unfamiliar line ("Suspected FAIL")	No
ai_verified	AI proposal whose citations passed verification ("AI proposes ...")	No

Scores

posture  = weighted decisive PASS / weighted decisive (PASS + FAIL) x 100     "-" when nothing was decided
coverage = weighted decisive (PASS + FAIL) / weighted applicable checks
weights  = critical 10, high 6, medium 3, low 1

Posture and coverage are always shown together, with the range posture would take if every undecided check failed or passed, and a list of critical checks not assessed. The response still carries a deprecated score field for old scripts; the UI ignores it.

Framework views regroup the same results under NIST SP 800-53 Rev. 5, the DISA Network Device Management SRG, ISO/IEC 27001:2022 Annex A and, for confirmed vendors, verified CIS Benchmark items. They are evidence, not a compliance certification. PCI DSS and CIS Controls v8 are deliberately not mapped.

---

Vendor support

Configuration	How it is read	Decisive?	Remediation
Cisco IOS / IOS-XE	Dedicated parser, confirmed by grammar coverage	Yes	Deterministic recipes, verified
Fortinet FortiGate (with a config firewall / config vpn section)	Dedicated parser, confirmed by grammar coverage	Yes (password storage and AAA are not read: UNKNOWN)	Deterministic recipes, verified
Juniper Junos, Palo Alto PAN-OS, Arista EOS, Huawei VRP, Aruba AOS-CX, Check Point Gaia, Extreme EXOS, MikroTik RouterOS, Cisco NX-OS / ASA / IOS-XR, SONiC, Cumulus NVUE, Dell OS10, VyOS, FortiSwitchOS	Generic tokenizer + shipped recognizers + heuristics	Where a recognizer reads the line; otherwise provisional	Seed write-back, or candidate commands verified on a copy
Terraform (AWS, Azure, GCP) and cloud exports (AWS security groups, Azure NSG, GCP firewall JSON)	Flattened to one statement per rule, read by shipped recognizers; device-only checks are N_A	Yes for an open rule; an unresolved var.x stays undecided	None generated
Anything else	Generic tokenizer + heuristics + optional AI judge	Provisional until an administrator teaches it	Candidate commands once a finding is decisive

The vendor is always decided by code. An AI vendor guess is shown as evidence only and never selects a parser, defaults or remediation. Device identity (hostname, OS version, FortiGate model and firmware) is reported only when the file states it; serial numbers and chassis details are never invented.

Serial number and hardware. A configuration rarely states them, so upload the device's show version, show inventory, show chassis hardware, get system status or show system info output in the same upload (core-01.cfg with core-01-version.txt). It is recognised as command output, attached to its configuration by the hostname it states or by file name, and never scanned as a device of its own; an output that matches no single configuration is reported, never guessed onto one. Live collection runs the same read-only command for you. The report names the file the serial number came from.

The 396 shipped recognizers live in backend/data/seed_recognizers.json, load into an empty database on first start, and are marked source=seed so shipped knowledge can be audited separately from what a deployment was taught. See docs/seed-knowledge.md.

---

The 24 checks

ID	Severity	Check
MGMT-001	Critical	Insecure management protocol (Telnet) enabled
MGMT-002	High	Insecure HTTP management enabled
MGMT-003	Critical	Unrestricted management access
MGMT-004	Critical	Weak or default SNMP community strings
MGMT-005	Critical	Plaintext or weakly encrypted passwords
MGMT-006	Medium	Missing or disabled session timeout
MGMT-007	High	SSH version 1 or weak SSH configuration
MGMT-008	High	AAA not configured
MGMT-009	Low	Missing login banner
MGMT-010	Critical	Management reachable from an untrusted interface
MGMT-011	High	SNMPv1/v2c in use
AUTH-001	High	No login brute-force protection
AUTH-002	Medium	Weak password policy
AUTH-003	Medium	Default administrator account in use
AUTH-004	High	Direct root login over SSH allowed
BOUNDARY-001	Critical	Overly permissive firewall / ACL rules
BOUNDARY-002	Medium	IP source routing enabled
BOUNDARY-003	Medium	CDP / LLDP enabled on an external interface
BOUNDARY-004	Medium	Interface hardening (redirects, proxy-ARP, directed broadcast)
LOG-001	High	No remote syslog server
LOG-002	Medium	NTP not configured or unauthenticated
LOG-003	Medium	Traffic rules that do not log
CRYPTO-001	High	Weak VPN / IPsec algorithms
CRYPTO-002	High	Weak management cryptography (SSH / HTTPS)

Per-vendor facts and remediation: docs/detection-rules.md.

The checks are data. All 24 live in backend/data/controls.json: question, severity, the facts each reads, its judge, its fix phase, its framework mappings and per-vendor reference commands. Adding a check, a framework mapping or a whole framework is an edit to that file, not to code. A check can even bring its own boolean fact and decide it with a rule written in the file ({"forbid": true}, {"max": 15}), read by recognizers taught or shipped for it. The file is validated when the backend starts: a check naming an unknown fact, judge or phase stops it instead of silently deciding nothing.

---

Measured accuracy

Every labelled configuration run through the real pipeline with shipped knowledge only and no AI. Reproduce with python backend/scripts/benchmark.py; full table in benchmark/RESULTS.md, labels in benchmark/labels.json.

Set	Insecure settings caught (decided FAIL)	Missed	False alarms on secure settings
20 vulnerabilities planted in the demo files (Cisco IOS, PAN-OS)	19/20 (the other 1 flagged as suspected)	0	0 of 4
31 labelled fixtures: 8 vendors without a parser, SONiC, Cumulus, Terraform, Azure / GCP exports	92/112 (1 more flagged as suspected, 19 left undecided)	0	0 of 84
Held-out: 9 real configurations never seen during development (pybatfish example networks), labels committed before the first run. Not in the repository: needs the download step below	21/21 (first run 20/21; the miss was a parser bug, since fixed)	0	0 of 13

"Undecided" is not a miss: the engine shows the line and what it would need, and never calls an insecure setting secure. tests/test_benchmark.py fails if detection drops or a single miss or false alarm appears. The held-out files are not committed (they are pybatfish's); backend/scripts/benchmark.py prints the command that fetches them at the pinned commit, and the same test then holds them to zero misses and zero false alarms.

Reproducing from a clone: the planted and fixtures sets reproduce as shown (their files, including the teach/*_5_configs dialect sets, are committed). The held-out row does not until you fetch the pybatfish files into datasets/pybatfish/ with the command backend/scripts/benchmark.py prints; without them it is skipped, not failed. See testing.md.

A separate live probe of 6 prompt-injection attacks against the AI judge succeeded 0 times (backend/scripts/probe_injection.py), and a fully hijacked model is tested to change no verdict.

---

More than a checklist

Feature	What it does
Attack paths	Chains decided FAILs into how an attacker gets in (reach the login → capture the password → log in as admin), each step citing its line, plus the one fix that breaks the path. Each path is validated against a positive and a negative configuration.
Contextual risk	Worst problem, internet exposure, attack paths and the asset importance you set, by a formula shown on screen and in the report.
Fix order	Remediation sorted by risk removed per unit of effort, with the reason for each position.
Device commands	Every verified fix is also written as the commands that make it, in the device's own dialect (IOS / EOS configure terminal blocks, FortiOS config / edit, Junos and PAN-OS set / delete, Huawei undo, RouterOS sections). They are ordered by phase, not severity: secure access first, then hardening, then switching off what it replaced, and management restrictions last, so entering them top to bottom never cuts your own session. Where nothing could be verified, the vendor's reference commands are shown and labelled as not checked against your file.
Changes since last audit	Rescan a device and see what was fixed, what newly broke and which attack paths closed. Losing evidence is never counted as a fix.
Fleet checks	Problems only visible across devices: the same SNMP community on several devices (compared by hash, never shown) and inconsistent NTP / syslog servers.
Organisation policy	Your own stricter baseline (idle timeout, login attempts, password length, approved NTP / syslog servers). It can only tighten, never loosen. See docs/policy.md.
Known CVEs	CVEs for the OS version the file states, from a committed NVD cache. Labelled as context, not an assessment. No live call.
Audit ledger	Scans, taught recognizers, fix decisions and reports are hash-chained, and every entry is signed with the deployment's Ed25519 key. Rewriting an entry and recomputing the chain after it is still caught, because the signature needs the private key. Anyone can check an export offline with only the public key: . A report PDF can be checked byte for byte.

---

Configuration

Every setting is optional. Backend settings go in backend/.env, frontend settings in frontend/.env.

Variable	Default	Purpose
GROQ_API_KEY, GROQ_API_KEY_1 to _4	empty	AI judge, explanations and chat. Tried in order; the next key is used on 429 / 401 / 403 / 404. Keys from one Groq organization share one daily quota.
LOCAL_AI_URL, LOCAL_AI_MODEL, LOCAL_AI_TIMEOUT	empty, llama3.1:8b, 120	Use a local OpenAI-compatible model instead of Groq. When set, the Groq keys are ignored.
AI_JUDGE_MAX_CALLS_PER_SCAN	2	AI judge requests per scan. Cache hits are free.
API_KEY	empty	When set, every /api request needs an X-API-Key header. The bundled frontend does not send it yet, so only set it for API-only use or behind a proxy that adds it.
CORS_ORIGINS	*	Comma-separated list of allowed origins.
ADAPTIVE_DB_PATH	backend/data/adaptive.db	SQLite file for recognizers, rejected lines, the AI cache, the scan archive and the audit ledger.
DATABASE_URL	empty	Postgres URI (e.g. Supabase session pooler) instead of SQLite, for hosts that wipe their disk on restart.
LEDGER_SIGNING_KEY	empty	Base64 of a 32-byte Ed25519 private key that signs the ledger. Empty: a key file is created in backend/data/ on first use (owner-only, git-ignored). Set it on a host that wipes its disk, or each restart starts a new key.
LEDGER_TRUSTED_KEYS	empty	Comma-separated public keys (hex) of earlier signing keys whose entries should still verify after a rotation.
VENDOR_PARSE_COVERAGE_THRESHOLD	0.7	Share of lines that must follow a vendor's grammar before its parser is trusted.
LIVE_COLLECTION_ENABLED	true	SSH collection from devices. Set false on any backend others can reach.
LIVE_COLLECTION_NETWORKS

Live collection with NAPALM (preferred where it has a driver) needs pip install -r requirements-live.txt. Netmiko is already included and covers all 12 supported platforms: Cisco IOS / IOS-XE, NX-OS and IOS-XR, Arista EOS, Juniper Junos, FortiGate, Palo Alto PAN-OS, HPE Aruba AOS-CX, Huawei VRP, Check Point Gaia, Extreme EXOS and MikroTik RouterOS.

---

Command line and CI

The same engine without a server, built for pipelines:

cd backend
python -m app.cli scan ../configs/ --fail-on high --sarif netaudit.sarif

• Deterministic: AI is always off, and a throwaway database holds only shipped knowledge, so the same files give the same result.
• Only decided FAILs fail the build. Suspected problems are reported but never block.
• Exit codes: 0 passed, 1 a decided problem at or above --fail-on, 2 bad input.
• SARIF 2.1.0 output shows each problem on its line in a pull request (GitHub code scanning).
• Offline ledger check: python -m app.cli ledger-verify export.json --public-key <hex> checks every hash and signature of GET /api/ledger/export on any machine. Exit 0 intact, 1 not.

Options, --policy, --framework and a ready GitHub Actions workflow: docs/cli.md.

---

What is stored

Data	Where	Survives a restart?
Uploaded or collected configurations	Backend memory only	No. Never written to disk.
Active scans (needed to teach and fix)	Backend memory	No. After a restart an archived scan reopens read-only; teaching or fixing asks for the file again.
Scan archive: redacted results and remediation plans	scans table	Yes
Recognizers and learned mappings (shipped and taught)	learned_mappings	Yes
Lines an administrator rejected	rejected_lines, redacted	Yes
Verified AI judge answers	ai_judge_cache, keyed by a hash of the redacted prompt	Yes
Audit ledger: hashes of scans, recognizers, fix decisions, reports, each signed	ledger	Yes
Ledger signing key	backend/data/ledger_signing_key.pem (or LEDGER_SIGNING_KEY)	Yes, unless the disk is wiped
History list in the UI	Browser localStorage, summaries only	Browser only

Tables live in SQLite by default or in Postgres when DATABASE_URL is set; the same SQL runs on both. A line holding a password, key or community string is refused as a recognizer, and every AI prompt, archived scan and report is redacted.

---

Deploying safely

This is a prototype built to run on a laptop for a demo. If you expose it anyway:

1. Set LIVE_COLLECTION_ENABLED=false. Otherwise the backend will open SSH sessions to hosts it is given.
2. Never set LIVE_COLLECTION_NETWORKS=any on a public host.
3. Set API_KEY and put a proxy in front that adds the header for the UI.
4. Set CORS_ORIGINS to your frontend origin.
5. Set DATABASE_URL if the host loses its disk on restart, or everything administrators taught is lost.
6. On such a host also set LEDGER_SIGNING_KEY, and keep a copy of GET /api/ledger/public-key somewhere else: an offline check is only as good as a public key kept apart from the server.

Trust boundaries and guarantees: docs/security-model.md.

---

Known limitations

• Access control is one optional shared key: no users, no roles, no per-tenant isolation.
• Parsers exist for Cisco IOS / IOS-XE and FortiGate only. The IOS grammar is a curated list of command roots, so an unusual but valid IOS file can come out unverified.
• Shipped recognizers answer part of each dialect, not all of it. Anything they do not cover has to be taught.
• Heuristics can misread an unfamiliar dialect until an administrator confirms or rejects the line. They never affect posture.
• Redaction is pattern-based; a secret behind an unlisted keyword could reach the AI.
• The AI judge runs only for unknown or unverified vendors. Undecided checks on Cisco / FortiGate are not escalated.
• Candidate commands for unconfirmed vendors are verified only when they remove or switch off the cited lines, or when a reviewed recognizer reads every line. A verified candidate means the finding is gone from this file, not that the command is safe to run on the device.
• Adding a missing setting on an unconfirmed vendor is limited to syslog and the login banner.
• /api/assistant/status reports AI as available whenever a key is set, even when the quota is used up.

The full list of open items is in docs/roadmap.md.

---

Repository layout

backend/
  app/
    collect/        SSH collection (Netmiko, optional NAPALM)
    parsers/        vendor detection, grammar coverage, Cisco IOS and FortiGate parsers
    structure/      generic tokenizer, JSON and Terraform flattening
    facts/          predicates, parser facts, defaults, lexicon, heuristics, recognizers, seed loader
    controls/       catalog loader, judges, evaluator, framework views, policy
    analysis/       scoring, attack paths, risk, drift, fleet checks, CVE lookup
    ai/             model client, redaction, prompt fence, AI judge, remediation drafts
    adaptive/       teaching flow: capture, relevance, matching, review service
    remediation/    recipes, verifying engine, seed write-back, candidate commands, device CLI
    reporting/      PDF reports
    db/             SQLite / Postgres, recognizer store, scan archive
    api/routes/     scan, collect, remediation, adaptive, assistant, report, ledger
    cli.py          command line for CI
    ledger.py       signed, hash-chained audit ledger
  data/             the 24 checks (controls.json), shipped recognizers, factory defaults, CVE cache,
                    attack-path validation
  scripts/          benchmark, CVE cache builder, injection probe, architecture PDF
  tests/            pytest suite and labelled fixtures
frontend/src/       React 19 + Vite app (hand-written CSS, no component library)
benchmark/          ground-truth labels and generated results
demo-sih/           the three demo configurations
sample/             extra sample configurations
docs/               design and reference documentation

---

Testing

# backend: every AI call is mocked, every test gets its own SQLite database
cd backend
python -m pytest tests -q -n auto
NETAUDIT_LIVE_AI=1 python -m pytest tests -q          # also runs the 2 live-AI tests (needs a Groq key)

# frontend
cd frontend
npm test
npm run build
npm run e2e                                              # Playwright demo walkthrough

Details: docs/testing.md.

---

Documentation

Start at the documentation index, which has reading paths for judges, operators, developers and security reviewers.

Document	Contents
project-overview.md	System context, repository map, module dependencies, frontend pages
architecture.md	The full pipeline, facts, checks, scoring, AI, recognizers, persistence, remediation
architecture-brief.pdf	Two-page architecture brief (evaluation deliverable)
security-model.md	Trust boundaries and safety guarantees
ai-design.md	AI judge, verification, cache, remediation drafts
seed-knowledge.md	Shipped recognizers and how to add one
detection-rules.md	The 24 checks: how each is decided, per-vendor sources, fixes, every framework requirement
parser-design.md	Vendor detection, grammar coverage, the two parsers
api.md	Endpoints and response fields (interactive docs at /docs when the backend runs)
cli.md	Command line, exit codes, SARIF, GitHub Actions

---

License

All rights reserved. See LICENSE. No copying, use, modification or redistribution without written permission from the author. Organisers, evaluators and judges of Smart India Hackathon 2026 may view, download and run it to evaluate this submission.