$ cat /forge/portfolio.md —— 12 flagship systems loaded
port | folio
Dragos Postelnicescu|AI Systems Engineer · 25+ years infrastructure

Everything here is real software I built and operate — designed, run, and repaired end-to-end on my own multi-node GPU platform, the Forge. This is my life’s work in the open: no notebooks, no toy demos, but autonomous agents that survive for days, an MCP platform with human-in-the-loop governance, a security appliance that reasons about my own network, and production pipelines with real daily users. Every system here earned its place by doing a job I actually needed done.

25+ yrsinfrastructure & telecom engineering
22+AI systems documented below
44MCP tools in production
OCIGenAI Professional 2025

Twelve systems that show how I work.

Each one solves a problem I actually had, runs on infrastructure I administer, and was reviewed by my own multi-model QA gate before it shipped. They range from battle-tested production systems to active research I’m still pushing on — the badges tell you which is which. Expand any card for the architecture and the measured results.

01 BridgeForgeProduction The platform that gives AI assistants governed hands on real infrastructure — auth, audit, approvals.

Letting an AI act on live systems safely is a governance problem before it is a tooling problem — and it’s the one I most wanted to get right. BridgeForge is my answer: an MCP server exposing 44 tools — shell, files, vision, knowledge base, task delegation, messaging — behind dual-token auth, with node enrollment across a genuinely heterogeneous fleet, a live observability dashboard, and an append-only audit trail where every action is attributable. High-risk moves don’t just execute; they stop at a tiered human-in-the-loop gate that messages me on Telegram and waits for an explicit yes. It’s the backbone the rest of the Forge runs on — including this very page, which was written, reviewed, and deployed through it.

> This portfolio was written, reviewed, and deployed through BridgeForge.

AI assistant MCP · dual-token auth 44 tools approval gate audit log
Measured results
  • 44 MCP tools live in production
  • Multi-node fleet enrolled — GPU servers → routers → a phone
  • Every action attributable in the audit log
MCPagent toolinghuman-in-the-loopobservabilityFastAPI
02 DesignForgeProduction A senior review panel as a service: design, code, spec, and structural critique on demand — or on every commit.

DesignForge is the most SaaS-shaped system in the fleet, and the one I lean on hardest day to day: an API-first review lane where every submission is grounded with its references inlined, routed by kind — design, code, spec, structural — to the model backends best suited to it, and returned as a structured verdict (APPROVE / REVISE with severity-ranked findings). Artifacts persist per run, and the orchestration layer calls it automatically for per-commit design-drift review, so nothing drifts quietly. The reason I trust it is simple: it has caught real defects in my own systems — floor overclaims, concurrency races, silent contract breaks — before they ever shipped. Having a tireless senior reviewer on call changes how you build.

submission + refs kind router critic panel structured verdict persisted artifacts
Measured results
  • 4 review kinds with automatic routing
  • Structured severity-ranked verdicts, persisted per run
  • Wired into per-commit design-drift review
LLM-as-judgeautomated code reviewstructured outputsCI-adjacentFastAPI
03 Ares / ContextForgeActive research A long-running agent that builds its own knowledge graph, audits itself, and recovers from its own failures.

Most things called “agents” are stateless request loops that forget everything the moment they answer. Ares is the opposite, and it’s the project closest to my curiosity about what a machine can become if you just let it keep running. A cloud model ensemble proposes, an adaptive weighting layer scores the votes, a judge model rules, and a stronger teacher model escalates the genuinely hard cases. What it learns is written into a persistent weighted memory graph — nodes, edges, trigger paths — that it maintains itself through its own auditing and consolidation loops, while an external watchdog enforces rate-limited revival whenever it falls over. It has run for days on end, grown its own knowledge, and picked itself back up without me — which is exactly the point.

model ensemble adaptive vote weighting judge teacher escalation memory graph
Measured results
  • Knowledge graph built and consolidated autonomously, with automated audit
  • 11+ days of continuous chained operation
  • Self-audit + consolidation loops live
autonomous agentsLLM orchestrationensemble votingvector memoryself-auditing
04 The Council + TrinityProduction Institutionalized disagreement: models that debate, rebut each other by name, and get judged — before decisions ship.

Single-model answers fail silently — they’re confidently wrong and nothing warns you — so I built the failure mode out. The Council seats five models as equals, local and frontier alike, through opening statements, named rebuttal rounds where they argue with each other by name, and a verdict from a fresh judge that never sat the debate, with an optional adversarial auditor and a toggleable Devil’s Advocate seat for when I want the case stress-tested. Trinity is its lighter, faster sibling: a bidirectional peer-review panel where every reviewer is themselves reviewed. Together they gate every v1.0 promotion across the fleet — and more than once they’ve caught me, and each other, mid-mistake. Disagreement, made into an institution.

R1 openings R2 cross-talk R3 fresh judge R4 adversarial audit
Measured results
  • 5-seat deliberation with cross-talk + fresh judge
  • Devil's-advocate seat toggles per call, no redeploy
  • Standing release gate for all systems
multi-agent systemsLLM evaluationAI governanceconsensus
05 GuardianForgeProduction A self-contained security analyst that watches my network, reasons about what it sees, and asks before it acts.

This is the system I’m proudest of as an end-to-end piece of engineering. GuardianForge bundles a network-model core, live radio and traffic sensors, and a knowledge store of every device and segment into a single deployable appliance. When something anomalous fires, it runs a real investigative cascade: a fast model drafts a verdict, a stronger tool-using investigator reads the draft, pulls more evidence from a curated tool registry — packet analysis, device history, evil-twin and radio-sibling checks — and a final pass writes an operator-grade incident report. An anti-hallucination gate refuses to emit any address that wasn’t in the evidence, and a three-tier enforcement ladder means dangerous actions are never automatic — they wait for my approval over Telegram.

sensors correlate incident draft → investigate → refine anti-hallucination gate tiered enforcement
Measured results
  • Full draft → investigate → refine cascade live on a hardened analysis node
  • Real incidents triaged end to end in ~80s, with human approval round-trips
  • Deployable to new hosts via the fleet’s enrollment cascade
security engineeringmulti-model cascadeanti-hallucinationhuman-in-the-loopanomaly detection
06 SVA · Screen VisionActive research A screen-vision agent that watches like an eye — cheap peripheral glances, expensive focus only where something moves.

SVA is my attempt to make machine vision economical the way biological vision is. Instead of running an expensive vision-language model over every frame, it captures screens over a live pipeline, builds a cheap peripheral change-map to see where something happened, and only then spends a deep read on the region that actually changed — a foveal attention pattern borrowed from how eyes work. It runs as a set of capture lanes with real-time change detection, feeding both interactive queries and batch deep-analysis. It’s wired into the platform as first-class vision tools, so any agent in the fleet can look at a screen and reason about it, on any enrolled machine.

screen capture peripheral change-map foveal deep read structured result
Measured results
  • Foveal capture lanes live in production as platform vision tools
  • Real-time change detection routing deep reads only where needed
  • Consumed by interactive queries and batch analysis across the fleet
computer visionVLM routingfoveal attentionreal-time systemsscreen understanding
07 ForgeNetworkProduction The always-on nervous system of the fleet — every node polled, measured, and charted, minute by minute.

Everything else on this page depends on the fleet being alive and observable, and ForgeNetwork is what makes it so. It runs as a daemon that polls every monitored node — routers, servers, links — on a tight cycle, collecting latency, reachability, WAN and VPN state, load, memory, connected clients, and periodic throughput tests, then stores it all as a time-series and surfaces it through REST endpoints and a live dashboard tab. It started as a single-router watchdog and grew, deliberately, into the monitoring backbone for the whole home-lab estate. When a node flaps at three in the morning, this is the system that already knows.

ping + SSH probe time-series store REST API live dashboard
Measured results
  • Polls the fleet on a 60-second cycle, 24/7, into a time-series DB
  • On-demand and scheduled throughput testing per node
  • Grew from one router to the whole estate as designed
network monitoringobservabilitytime-seriessystemd daemoninfrastructure
08 TranslateForgeProduction An EN→RO literary pipeline that a professional translator uses on real books, every working day.

Literary translation is where ordinary machine translation goes to die: voice, idiom, and consistency across a whole book matter far more than getting the gist right. TranslateForge is a multi-pass LLM pipeline with a reviewer-in-the-loop UI, built for — and continuously shaped by — a working professional literary translator who uses it on real manuscripts every working day. Every one of its eleven-plus sprints of iteration came from an actual book meeting an actual deadline, not a synthetic benchmark. It’s the clearest proof I have that these systems earn their keep only when someone genuinely depends on them, and lives on its own public subdomain because it’s a real product, not a demo.

manuscript multi-pass pipeline consistency layer reviewer-in-the-loop UI
Measured results
  • In daily professional use
  • 11+ sprints driven by real-work feedback
  • Deployed on its own public subdomain
applied NLPLLM pipelineshuman-in-the-loopproduct iteration
09 TradeArchitectResearch / live monitoring Model championships for markets: measured win rates, consensus architectures, portfolio-level risk control.

Every model earns its seat here the hard way: per-timeframe champions are selected by measured evaluation win rate, and the top timeframe runs a live consensus of the local champion, an API reasoning model, and a frontier judge before it will commit to a call. A real-time futures monitor feeds the whole thing, and a portfolio-level liquidation-surface controller manages cross-margin risk across positions. I want to be precise about what this is: all figures here are measured evaluation metrics from the harness, and this is research instrumentation — a study in whether disciplined model competition and consensus can be quantified — not a returns claim or financial advice. The discipline of measuring honestly is the actual product.

live market feed champion selection 3-model consensus risk controller
Measured results
  • M15 champion measured at 75.8% eval win rate
  • H1 consensus architecture in validation
  • Portfolio-level liquidation management built
quantitative researchensemble modelsreal-time datarisk management
10 ChessForgeResearch A chess agent that doesn’t learn weights — it evolves the English prompt that describes how to think.

ChessForge is where a lot of the Forge’s honesty machinery was first forged in fire. Rather than train a network, it evolves the natural-language reasoning prompt itself — the plain-English description of how to think about a position — and tests each generation against a real Stockfish opponent for an honest signal instead of self-play echo chambers. Tactics it discovers are promoted into a knowledge store with confidence scores; ones that stop working are demoted. The hard lesson that shaped everything downstream: it once graded itself brilliantly while losing every game, so I built strict anti-hallucination checks and grade caps that refuse to reward a system with zero wins. Fully explainable, and honest to a fault by design.

reasoning prompt play Stockfish verify tactics promote / demote evolve prompt
Measured results
  • Self-evolving natural-language strategy, no weight training
  • Honest signal from real Stockfish games, not self-play
  • Anti-hallucination + zero-win grade caps born here, reused fleet-wide
prompt evolutionexplainable AIself-improvementvector memoryhonest evaluation
11 The BrainActive research · early An orchestration layer that lets even small local models drive the whole stack — safely, through a single gate.

The Brain is the next step in the whole program, and the one I’m building most carefully. It sits above the review and deliberation stack and asks a hard question: how do you let a weak, cheap local model — or a frontier one, or me by hand — drive powerful capabilities without anything unsafe slipping through? The answer is a single non-bypassable dispatch chokepoint that mints a signed ticket for every action; organs refuse to move without one. Its governing doctrine is that the digest informs and the grammar constrains — an unsafe action is simply absent from the menu a small model is offered, never a rule it has to remember to follow. It carries layered memory, a self-repair engine, and an ownership matrix ratified through my own Council. Early, unfinished, and the piece of architecture I’m most excited about.

any driver model single dispatch gate signed ticket constrained action menu self-repair
Where it stands
  • Single non-bypassable dispatch chokepoint live, ticket-gated
  • Drivable by weak local models and frontier models alike
  • Ownership matrix Council-ratified; verification suite green
agent orchestrationcapability safetylayered memoryself-repairgrammar-constrained
12 Forge ArchitectsProduction + research A family of “architect” services that route work to the right model, judge the output, and learn from every correction.

Across the fleet a single pattern kept proving itself, so I made it a family. Each Architect owns one domain — vision, translation, prompts — and does the same disciplined job: resolve the best prompt for the task, route to the right model backend, judge the result on explicit quality axes (accuracy, completeness, structure, specificity, no-hallucination), and fold every human correction back into a learning store so the next answer is better. TranslateArchitect runs in production inside TranslateForge, learning from a professional’s edits; VisionArchitect governs every vision workload with a routing table and a five-axis judge; PromptArchitect underpins the agent pipelines. It’s the closest thing the Forge has to a house philosophy, written as running code: never trust one model blindly, always judge, always learn.

route resolve prompt dispatch model 5-axis judge learn from corrections
Measured results
  • TranslateArchitect in daily production, learning from a professional’s edits
  • VisionArchitect governs all vision workloads — routing + 5-axis judge
  • Shared corrections-learning loop backed by vector memory
LLM routingLLM-as-judgequality governancecontinual learningreusable architecture

Three technologies hold it all together.

The flagships are not islands. They stand on three shared foundations — an engine, an optimizer, and a memory — that I built once and reuse everywhere.

CloudMind
The multi-model consensus engine the fleet is built on — the shared brain behind every place two models are asked to agree.
DTSE+GA
A proprietary adaptive optimization layer, so the stack tunes its own runtime parameters and timeouts instead of relying on brittle hand-set constants. The system learns its own settings.
ForgeDossier
A ChromaDB-backed institutional memory that every agent reads and writes — so the fleet remembers its own decisions, and no hard-won lesson is lost between sessions.

The Forge: research-grade AI on production-grade operations.

Everything above runs on a self-hosted, multi-node GPU platform I designed and administer end-to-end: built primarily in Python (FastAPI), service management under systemd, public edges through Cloudflare tunnels, a strict credential-handling pattern, tiered verification gates (L1–L3) before anything is promoted, automated cross-node backups, and institutional memory so no decision is lost between sessions.

> I run AI systems the way production infrastructure is run — because that is where I come from.

6+
nodes in the fleet
25+
services under management
L1L3
verification gates
24/7
operation

And the rest of the fleet.

Beyond the twelve flagships — the working systems that fill out the platform. Smaller in scope, but each one real, each one built to do a specific job well.

EarForge
A live audio copilot that listens to a conversation or stream in real time and keeps useful context flowing alongside it. Built for the moments when reading a transcript after the fact is already too late.
MailForge
A multi-identity mail aggregator that pulls several inboxes into one place and puts task-specific agents to work on them — triaging, summarizing, and surfacing what actually needs a human. Email as a data source, not a chore.
PDForge
A PDF extraction pipeline that turns messy real-world documents — scans, forms, mixed layouts — into clean structured text the rest of the fleet can reason over. The unglamorous plumbing that a lot of other work quietly depends on.
SpeculativeOracle
A conditional-futures prediction platform: given a set of premises, it reasons about what follows and how confident that inference really is. An experiment in making models state their assumptions out loud.
CamForge / CollageForge
Multi-camera pipelines for a fleet of IP cameras, with hardware-accelerated NVENC compositing into live HLS mosaics. Several streams, one wall, encoded on the GPU rather than melting the CPU.
CloseUI + MCR
A fork of a popular open chat UI (“Open” → “Close”) paired with MCR, a Model Capability Router that sits between the UI and every LLM backend to unify failover, task-based routing, and capabilities. The base is built and running on the Forge; the full revival is on the roadmap, with The Brain slated to register as a backend.
ForgeSpec./visit
Specification-as-a-service — turning loose intent into rigorous, structured specs, with its own public presence at forgespec.dev. The discipline of writing things down properly, made into a tool.
Math Agent
A self-teaching mathematics prover that works through problems and builds up its own capability rather than being handed answers. An ongoing research experiment, intended to eventually graft into the autonomous-agent line.
CloudMind v2
The multi-model consensus engine that underpins much of the fleet, currently getting a ground-up makeover into a cleaner second-generation chatbot. The shared brain a lot of these systems quietly call into.

Who runs this.

AI Systems Engineer — 25+ years across enterprise infrastructure and telecom, most recently as a Principal Technical Support Engineer. MSc Open Systems Architecture and BSc Electronics & Computer Science, Politehnica University of Bucharest. Oracle Cloud Infrastructure 2025 Generative AI Professional and AI Foundations Associate.

Open to senior AI/infrastructure engineering roles — full-time or B2B contract.

>Bucharest, EU >25+ yrs infrastructure >OCI GenAI Professional '25 >Full-time or B2B contract

Talk to me.

Email is the front door — or use the form on the main site. Prefer voice? Request a call by email and I'll confirm a slot.

./email me ./contact form ./linkedin
Calls
By appointment — request via email
Base
Bucharest, EU · remote-ready