π§ Contents β the page is long on purpose; jump straight to what you need
- About me
- Operating principles
- Expertise matrix
- AI & LLM engineering β RAGRAF Β· AI-Consensus Β· ctf Β· LLM Code Projector Β· MD.Viewer Β· sr-search-replace Β· MCP reference
- Infrastructure engineering β Windows & RDP Β· email Β· web/PHP Β· Linux
- Tech stack
- Live metrics
- Quality bar
- Documentation & reference
- Repository navigator (collapsed catalogue below)
- Collaboration Β· Connect
I am Mikhail Deynekin β a senior sysadmin and developer based in Frankfurt am Main, Germany, in IT continuously since 1991: from DOS workstations and BBS boards to Kubernetes-era infrastructure and, today, LLM agent tooling. I build the kind of software I want to operate at 3 a.m.: idempotent, self-documenting, honest about its failures, and measurable.
- π§ 35+ years across Windows & Linux infrastructure: automation, hardening, forensics
- π€ Current focus: code knowledge graphs for agents (RAGRAF), MCP contracts, multi-model consensus (AI-Consensus)
- π§ Email infrastructure specialist: hMailServer, Exim, Roundcube, SpamAssassin, SPF/DKIM/DMARC
- π₯οΈ RDP security toolkit author and the Windows 11 25H2 enablement script used in production fleets
- π§Ύ Docs-as-product: every serious repo ships architecture notes, agent regimens, options references
- π deynekin.com Β· π¬ Mikhail@Deynekin.com Β· π replies within 24 h
Note
How I work in one sentence: measure first, automate second, document always β and never let a tool guess silently when it could say βI donβt know, here is why.β
| # | Principle | What it looks like in code |
|---|---|---|
| 1 | Verifiable facts over plausible text | RAGRAF answers only with file:line proofs and provenance: stated/derived; benchmarks live in bench/results/, not in marketing |
| 2 | Honest degradation, never silent failure | No Joern layer β capability=degraded and a plain refusal from slice; Clear-TempFolders ships a DryRun mode; ctf never executes what it reads |
| 3 | Idempotence by default | hMailServer-Spam-Provisioner is safe to re-run on every boot; ctf snapshots are reproducible with per-file hashes |
| 4 | Zero-dependency core, optional power layers | ctf = one POSIX script; RAGRAF embeddings run without torch, Joern is optional; MD.Viewer needs no build, DB, or framework |
| 5 | Documentation is the product | Six-part council design in AI-Consensus, an agent integration regimen in RAGRAF, token-budget guides in ctf β and this profile |
| Domain | Depth | Evidence in this profile | Core tech |
|---|---|---|---|
| AI & LLM engineering | Building agent tooling: RAG, MCP servers, multi-model consensus, context pipelines | RAGRAF Β· AI-Consensus Β· ctf Β· LLM Code Projector | Python 3.13, PHP 8.4, SQLite (graph+FTS5+vectors), tree-sitter, MCP |
| Email infrastructure | Production mail since the pre-cloud era: MX, filtering, deliverability, webmail | 7 repos under Email infrastructure | hMailServer, Exim, SpamAssassin, Roundcube, SPF/DKIM/DMARC |
| Windows & RDP security | Fleet automation, session forensics, upgrade engineering, security-log analysis | 12 repos under Windows & RDP security | PowerShell 5.1β7.6, C#, Windows Security log, BITS, GPO-level hygiene |
| Web & PHP backend | Single-file engines, IIS + FastCGI deployments, WordPress/WooCommerce ops | HTML.PHP Β· MD.Viewer Β· IIS-PHP-FastCGI-AutoDeploy | PHP 8.3+, IIS, WordPress, WooCommerce |
| Databases | PostgreSQL/MySQL operations: encoding migrations, cache truncation, tuning & replication | pg-encoding-migrator Β· Clean-RoundCube-Cache-PostgreSQL-Windows | PostgreSQL, MySQL/MariaDB, PDO with prepared statements |
| Linux & network ops | RouterOS filtering, certificate hygiene, disk/power management, cron discipline | sata-spindown-guard Β· cert-unifier Β· Bash_WP-CLI_Update | Bash, RouterOS, hdparm, Letβs Encrypt |
Not a pile of experiments β a connected toolchain. Each tool closes one gap of the agent-assisted development loop, and every gap I hit in my own work became a repo:
Answers βwhere is it, who calls it, what breaks if I change it, how is it done elsewhereβ with verifiable
file:linefacts β instead of blindrgand full-file reads. One SQLite per corpus holds the graph, the BM25 lexicon, the vectors and the drift ledger.
Indexing pipeline & stores
Measured, not promised β real PHP portal, 14 navigation tasks, artifacts in bench/results/:
| Metric | Without the index | With RAGRAF |
|---|---|---|
| Tool calls for 14 tasks | 28 (modelled realistically 43) | 14 |
| Average calls per task | 2.0 / 3.1 | 1.0 |
| Candidates to review | up to 34 files | 1β5 exact |
| Files read in full | 14 (minimum) | 0 |
| Target-file hit on the first query | β | 92 % (12/13) |
Verifiability (facts with file:line) |
β | 100 % |
Three design decisions, each born from a measured defect
- Immutable index + drift ledger instead of βupdate on the flyβ. Inside a revision the index never changes; code edits accumulate in the ledger and affected facts turn
stale. Mixing timestamps inside one answer is exactly how wrong conclusions are born. - No free-form graph query language for the agent. Fixed operations only: in LLMxCPG (USENIX Security β25) base models produced 132/1278 and 19/1278 valid CPGQL queries versus 1278/1278 for a fine-tuned one β a query language turns base models into error generators.
- Every fact carries proof:
path:line,provenance(stated/derived),index_rev,confidence. An edge without evidence looks credible only because it came βfrom the databaseβ.
π MCP tool reference β the core ragraf_* contract (23 of the 59 tools in RAGRAF v0.3.0)
| Tool | Answers |
|---|---|
ensure |
bootstrap: corpus is indexed and current before the session starts |
index_status |
revision, parse completeness, degradation, ledger size |
drift_report |
ledger of divergences between index and code |
drift_ack |
acknowledge ledger entries after review |
search |
βwhere is this written aboutβ β hybrid BM25 + vector; --with-text returns snippets from the index |
symbol |
definition site: kind, file, lines, signature |
references |
where else the symbol is used |
callers |
who calls it |
callees |
what it calls |
impact |
what a change affects β proven and heuristic links reported separately; mandatory before edits |
slice |
structural slice (nodes, edges, lines) without function bodies β full capability only |
code |
symbol bodies; text from the index by default, `text_source: index |
repo_map |
repository map within a character budget; chars_used is measured, not estimated |
schema |
what is in a DB table (from migrations) |
table_usage |
who reads or writes the table |
routes |
routes and entry points |
live_read |
the current file on disk with differences from the index highlighted |
policy_trace |
who gets a permission, and under which condition |
class_usage |
a styling class: where defined, where applied, what diverges |
write_check |
is this write valid β write plan checked against the schema before running |
usage |
call telemetry and effect estimation |
freshness |
is the index current (checksum comparison) |
catalog |
corpus registry: what is indexed, with which checksums |
There is, and will be, no free-form graph query tool (FT-15): a MATCH β¦ RETURN
attempt is rejected with code -32602.
Quick start
ragraf init <project-dir> # creates config/myproject.yaml
python tools/ragraf.py --project config/myproject.yaml index{
"mcpServers": {
"ragraf": {
"command": "python",
"args": ["<repo>/tools/ragraf-mcp.py", "--project", "<repo>/config/myproject.yaml"]
}
}
}Important
Degradation is a feature. No Joern layer β capability=degraded and an honest
refusal from slice; no embedding model β lexical search still works; no free port β
the server says why. Silent degradation never happens.
A PHP 8.4 web application where several models deliberate in defined roles, a consensus engine scores their agreement, and every answer ships a clickable DAG of the whole session. Full audit trail: every model response, every step, every score β in MySQL.
- SKILLS system β declarative JSON workflows defining agent roles, prompts and consensus rules; the router auto-selects or auto-creates the most relevant SKILL per query.
- Strict PHP 8.4 β strict types, readonly properties, enums throughout; six-part design doc from architecture to the full CouncilEngine module.
Tip
The council pattern shines where single-model answers are expensive to trust: research synthesis, policy-style questions, architecture reviews.
One self-contained POSIX script (bash 4.2+, shellcheck-clean) that folds a source tree into a single Markdown / JSON / JSONL / text document with relative paths β the document an LLM reasons over best, or a reviewer reads in one pass.
ctf --budget 60000 src/ > context.md # LLM context, capped at a token budget
ctf --format jsonl --hashes . > audit.jsonl # auditable snapshot, per-file hashes| Format | Best for |
|---|---|
| Markdown | LLM context & cross-file review: syntax-highlighted, path-annotated |
| JSON / JSONL | machine pipelines, chunkers, RAG ingest |
| Text | diffs, greps, plain archives |
It does not shell out to eval, does not execute anything it reads, never writes
outside the file you asked for.
The reverse direction of the loop: feed it a ChatGPT/Claude/DeepSeek/Gemini answer with project trees and code fences β get a runnable directory tree in one command. Recognizes ASCII and Unicode trees, any nesting depth.
Note
The repository is archived: the released v6.0.1 tool keeps working and stays installable, and the code is kept public for reference without further updates.
./MD_to_Files.sh chatgpt_response.md ./my-new-projectThe loopβs last mile: LLM- and human-written markdown rendered as polished docs.
Auto TOC with heading numbering, Mermaid with auto-repair, glossary tooltips, dark mode,
code-copy controls, searchable file browser, built-in self-updater β no build step,
no database, no framework: copy md.php next to your .md files and open it.
Universal search & replace with session rollback, binary detection and backup management: the safe way to let an agent (or yourself) touch forty files at once.
Warning
Agent regimen I actually follow (from RAGRAFβs integration doc): start with
index_status + drift_report; never base a conclusion on a fact marked stale β
live_read it first; whole-file reads are the last resort, not the first reflex.
The other half of my life: systems that must stay up, stay secure, and stay boring.
| Project | Role |
|---|---|
| Windows-11-25H2-Update-Script | 24H2 β 25H2 via official enablement package KB5054156: BITS fallback, retry logic, signature validation, full logging |
| RDP-Diagnostic-Tool | Complete RDP readiness diagnostics with auto-remediation & reporting |
| 1st-Remote-Session-Manager-Pro | Shadow, view-only, disconnect, message, logoff β with GitHub auto-updates |
| RDPAudit | Real-time security monitoring & session auditing (C#) |
| 1st-RDPMon | Server-side session monitoring & alerting |
| Get-Windows-Security-Events-By-IP | Forensic Security-log analysis by IPv4/IPv6 with failure-code decoding |
.\Upgrade-Win11-To-25H2.ps1 -RetryCount 5 # enterprise guardrails baked in
- hMailServer at the core: domains, accounts, aliases, lists, CSV/JSON export, antivirus integration (Defender, Kaspersky, ClamAV)
- SpamAssassin as a Windows service under PowerShell 7 with Task-Scheduler rule updates
- Deliverability stack: SPF Β· DKIM Β· DMARC configured and monitored, not βset and forgottenβ
- Roundcube ops: deployment on IIS + FastCGI, PostgreSQL backend, cache truncation tooling
Tip
Mail troubleshooting order I trust: DNS & rDNS β SPF/DKIM/DMARC alignment β spam-score breakdown β content rules. Nine times out of ten the problem is the first two.
- HTML.PHP β single-file PHP 8.3+ template engine: Apache SSI directives,
html4.pl-compatible API, XSS-safe by default - IIS-PHP-FastCGI-AutoDeploy β one-shot IIS + PHP FastCGI installer (Roundcube-ready)
- pg-encoding-migrator β safe PostgreSQL encoding migration (e.g. WIN1251 β UTF-8)
- woocommerce-1c-data-cleaner Β· Bash_WP-CLI_Update Β· Clean_BOM_Senior Β· BAS-IP
- sata-spindown-guard β cron-based HDD spindown guard (
hdparm), safe for unmounted drives - cert-unifier β unified PEM (key + chain) for Letβs Encrypt deployments
- RouterOS filtering & monitoring practice; bash tooling with cron discipline and dry-run habits
Snapshot of 2026-11-07: 48 product repos Β· 40.9 MB of code Β· 20 languages Β· 54 stars (public, non-fork; this profile repository excluded).
The two cards are local SVG snapshots rebuilt from the GitHub API with
python tools/build_cards.py --write β immune to third-party rate limits and
outages; the only live third-party card is streak-stats.
| Signal | Where to verify |
|---|---|
| shellcheck-clean tooling | ctf v4.1.0 README badge & CI-equivalent checks |
| SemVer discipline | ctf v4.1.0 Β· Projector v6.0.1 Β· MD.Viewer v3.0.1 (md.php v2.18.0) Β· AI-Consensus v1.0.0 |
| Honest licensing | PolyForm Noncommercial 1.0.0 in 32 repos Β· BUSL-1.1 in RAGRAF and RDPAudit Community Β· MIT in 14 utilities and package formulae β a LICENSE file in every repository |
| Self-updating deployments | MD.Viewer updater.php v3.14.0 Β· RSM-Pro GitHub auto-updates |
| Measured performance | RAGRAF bench/results/ artifacts Β· ctf Performance section |
| Security defaults | signature validation (Win11 script) Β· XSS-safe escaping (HTML.PHP) Β· admin checks (Print-Spooler-Queue-Cleanup) Β· prepared statements (AI-Consensus) |
| Token-budget awareness | ctf --budget Β· ragraf_repo_map chars_used measured volume |
Every serious repo ships its own field guide. The most useful entry points:
| Document | Where | What it teaches |
|---|---|---|
| Agent integration regimen | RAGRAF β docs/implementation/INTEGRATION.md |
Call order, drift discipline, hard rules, MCP config β copy-pasteable into AGENTS.md |
| Capabilities & design decisions | RAGRAF β README |
The fixed-operation contract (59 MCP tools in v0.3.0), 29 tree-parsed languages, three decisions with their measured defects |
| Council engine design, parts 1β6 | AI-Consensus β docs/ |
Architecture, SKILL JSON schema, MySQL DDL, controller flow, full PHP 8.4 engine |
| Options reference & token budgets | ctf-collect-to-file β README |
Every flag, four formats, budget fitting, exit codes, Windows notes, performance |
| Input format & instructions for LLMs | LLM Code Projector β README |
How to format LLM answers so they materialize into clean file trees |
| Zero-build docs deployment | MD.Viewer β README |
md.php + updater: self-hosted documentation without a toolchain |
| Enterprise upgrade quickstart | Windows-11-25H2-Update-Script β README |
PowerShell & Batch entry points, parameters, logging layout |
| Idempotent provisioning contract | hMailServer-Spam-Provisioner β README |
What re-running changes (nothing) and guarantees (everything) |
| Service-mode SpamAssassin on Windows | SpamAssassin-Windows-PS7 β README |
spamd as a system service + daily rule updates via Task Scheduler |
ποΈ Repository navigator β all 48 product repos by domain
AI & LLM tooling RAGRAF Β· AI-Consensus Β· RevoAp Β· ctf-collect-to-file Β· LLM Code Projector Β· sr-search-replace Β· github-agent Β· github-wiki-llm-writer Β· agent-prompt-kit
Windows & RDP security Windows-11-25H2-Update-Script Β· RDP-Diagnostic-Tool Β· 1st-Remote-Session-Manager-Pro Β· 1st-RDPMon Β· RDPAudit Β· Get-Windows-Security-Events-By-IP Β· Windows-LogOff Β· windows-show-all-tray-icons Β· Windows-Cleaner-and-Optimizer Β· Clear-TempFolders Β· Print-Spooler-Queue-Cleanup Β· Import-CaCertificates
Email infrastructure hMailServer-Spam-Provisioner Β· SpamAssassin-Windows-PS7 Β· PowerShell_Export-hMailServerData Β· hMailServer-External-Antivirus Β· hmailserver-kaspersky-cleanup Β· hmailserver-clamav-cleanup Β· Clean-RoundCube-Cache-PostgreSQL-Windows
Web, PHP & databases HTML.PHP Β· IIS-PHP-FastCGI-AutoDeploy Β· Bash_WP-CLI_Update Β· woocommerce-1c-data-cleaner Β· BAS-IP Β· Clean_BOM_Senior Β· pg-encoding-migrator Β· regru-control Β· fornex-control
Linux, network & security sata-spindown-guard Β· cert-unifier Β· mikrotik-doh-failover Β· CVE-2026-111497
Spam & abuse defence Contact-Form-7-LLM-Spam-Checker Β· PHP-Form-Spam-Checker Β· Fail2Ban_AbuseIPDB
Documentation & content MD.Viewer Β· MD.Scribe
Packaging homebrew-github-agent-cli Β· scoop-github-agent-cli
π¬ Email infrastructure arsenal β extended
- hMailServer β domains, accounts, aliases, distribution lists, CSV/JSON export, antivirus integration (Defender, Kaspersky, ClamAV)
- SpamAssassin on Windows β
spamdas a system service under PowerShell 7 with daily rule updates via Task Scheduler - Idempotent provisioning β
[SPAM]IMAP folders + filter rules for every account, safe to re-run - Virus cleanup β Kaspersky and ClamAV scan-and-quarantine tooling for the hMailServer store
- Deliverability stack β SPF Β· DKIM Β· DMARC configuration and hygiene
- Roundcube β deployment on IIS + PHP FastCGI, PostgreSQL backend, cache truncation tooling
- Exim β routing, filtering, delivery tuning on Linux
π₯οΈ RDP & remote access toolkit β extended
- RDP-Diagnostic-Tool β full readiness check & auto-remediation
- 1st-RDPMon β real-time server-side session monitoring & alerting
- 1st-Remote-Session-Manager-Pro β shadow, control, disconnect, message, logoff
- RDPAudit β security monitoring & session auditing (C#)
- Get-Windows-Security-Events-By-IP β brute-force & auth-failure forensics
- Windows-LogOff β forcible logoff with residual-process cleanup
πͺ Windows automation & security β extended
- Windows-11-25H2-Update-Script β enablement-package upgrade with BITS fallback & signature validation
- Windows-Cleaner-and-Optimizer β safe browser/system cache cleanup
- Clear-TempFolders β TEMP hygiene with DryRun mode
- Print-Spooler-Queue-Cleanup β forcible spooler queue clear with admin checks
- Import-CaCertificates β CA root/intermediate import into Windows stores
- windows-show-all-tray-icons β notification-area registry tweak
Where I add value fastest:
- π§ Email deliverability rescues β SPF/DKIM/DMARC alignment, spam-filter tuning, hMailServer/Exim migrations
- π₯οΈ RDP hardening & session operations for Windows fleets, from diagnostics to forensics
- πͺ Windows upgrade/update automation with enterprise guardrails (signatures, retries, logging)
- π€ LLM agent tooling β code knowledge bases, MCP contracts, context pipelines, consensus patterns
- π§Ύ Evidence-based codebase audits β RAGRAF-driven, every finding with
file:line
Open-source collaboration is welcome on any repo: issues with reproduction steps get answers, pull requests with tests get merged.
- π Building: RAGRAF β making code knowledge bases a first-class citizen for LLM agents
- π¦ Recently shipped: RAGRAF v0.3.0 Β· github-agent v1.3.0 Β· MD.Viewer v3.0.1 Β· cert-unifier v3.1.1
- π± Exploring: agent-safe MCP contracts, consensus scoring for multi-model councils
- π― Open to: collaborations on AI tooling Β· consulting on email & RDP infrastructure
- π« Reach me: Mikhail@Deynekin.com β I reply within 24 hours