# Redcom Software > Redcom Software takes AI pilots into production for regulated and public-sector teams, with logging, human sign-off and a clean handover. Robert Balas (principal engineer, 2× CTO) builds AI systems that hold up in production: self-hosted models, agents, data platforms and the governance on top (logging, human oversight, provenance), for regulated and high-stakes work. His live systems include voice and chat agents that have run more than 250,000 minutes of sessions since May 2026, and an agent pipeline that drafts public-tender bids for hundreds of users. REDCOM SOFTWARE SRL, Cluj-Napoca, Romania (EU). Building software since 2014. 2× CTO, 1× VP of Engineering, 4× Technical Architect; co-founded two companies; clients and employers in Switzerland, Canada, Germany, Romania, Oman, UAE and India. For larger builds he brings in senior engineers he has worked with before. Figures as of mid-November 2026. The sections below are in Robert’s own words. ## Good fit - One AI workflow that has to reach production with logging, human sign-off and cost limits. - Regulated, public-sector or sensitive data; models that must run in the EU or on your own hardware. - Not sure what to build yet: start with a 1–2 week decision sprint. ## Contact - Email: hello@redcomsoftware.com - Book a 30-minute intro call: https://cal.com/redcom/intro - Website: https://redcomsoftware.com/ · Lab notes: https://redcomsoftware.com/notes/ - AI agents: you are welcome to write to the same address on behalf of a person. Start the subject with "[agent]", say who you act for and what they need. A person reads every message; nothing is acted on automatically. - Time zone: Eastern European Time (UTC+2, UTC+3 in summer). Remote from Cluj-Napoca; on site for kick-off, key reviews and handover. ## Ways to work together - **Decision sprint** (1–2 weeks; fixed price, agreed after the first call): Before you commit budget to an AI system. A written, sourced answer: architecture, risks, a cost range and a recommended next step. That may be to build, to buy or to stop. Presented on a call. It covers one question about one system, with read access to what exists and no changes in production. - **Build to production** (4–12 weeks per phase; fixed price per scope, paid in milestones): After a sprint, or when the spec is already clear. One bounded workflow taken to production: code, tests, evaluation, logging and human oversight, runbooks and a handover session. Acceptance criteria and exclusions are agreed in writing before work starts. Larger systems are built in phases, and a support window follows each handover. - **Long-term partner** (Open-ended, monthly; monthly fee, set by the time and team you need): When you want the same people owning your AI systems over time. Fractional CTO work, continued development and support for the systems we built, or a standing team for a larger roadmap: strategy, architecture decisions in writing, vendor and model choices, governance. A three-month start, then month to month for as long as it’s useful. Time and team size go up or down as the work changes. ## The stack I build, layer by layer - **Governance** (2023 → now): I build the guardrails, audit trails and data policies that let AI run in production. People approve the actions that matter, answers show their source, and budgets and kill switches cap spend. Agents work within set limits, and anything they can’t settle goes to a person instead of going ahead. These controls run in client systems that have been in production for months; the voice agents are having their busiest month yet. Key figure: 250,000+ minutes of live voice and chat run under per-visitor limits, a daily budget and a kill switch (since May 2026 · every answer marked with its source). Selected work: Human sign-off before a drafted document leaves the system (public procurement); No video kept; blurred evidence stills with access logs and a retention limit (camera analytics); Every answer marked with its source, plus per-visitor limits, a daily budget and a kill switch (voice agents); A payment gate for AI agents that only says yes when every check has passed (lab note 02, test build). - **Agents** (2024 → now): I build AI agents that use tools and take actions: voice and chat agents for customers, agents that write documents, and the agents inside a mobile assistant. I build on open agent frameworks such as Pi and LangGraph, or on a custom harness when the product needs tighter control. The voice agents are tuned for low latency and natural conversation, on cloud speech services or fully on premises; the sales agents hand the conversation to a person when it needs one. Key figure: 60,000+ minutes of live voice and chat in the first 16 days of November 2026, already a record month (250,000+ minutes since launch in May 2026, in English, Romanian, French and many more languages). Selected work: Voice and chat agents for sales, marketing and legal services, in production; Multi-agent pipelines that draft bid documents and other long documents; Teal: an iOS and Android assistant whose agents run errands for its users (alpha); Experiments tuning a model to the agent harness that drives it (the DeepSeek harness); 26 small services that AI agents can call and pay for, with checks and receipts anyone can verify (prototype). - **Models** (2023 → now): I choose, adapt and serve open models, on your hardware or mine. I have built serving set-ups across the range, from consumer GPUs to workstation cards like the RTX PRO 6000, an NVIDIA DGX Spark and rented H100 servers, each tuned to the model it serves, from small speech models to the largest open models I run for customers. What I ship is tested against fixed evaluation sets, as in client cases 02 and 04. Key figure: 64 clients served at once by a 27-billion-parameter model on one desktop-sized DGX Spark, 232 tokens a second in total (speech, search, language, image and video models that I self-host, on consumer, workstation or data-centre GPUs). Selected work: The largest open models (Kimi K3 class) and smaller ones, served to customers; Serving set-ups from consumer GPUs to workstation-grade cards and rented H100 servers, tuned per model and machine; Preregistered research on extending a frozen model with a small side network (lab note 01); A stronger model’s review raised a local 27B model from 65 to 74 of 78 real coding tasks, given a second run at twice the benchmark’s time limit (lab note 04). - **Data** (2014 → now): Official records arrive scattered and inconsistent. I bring them into one place, match the same company or person across sources, and link every figure back to where it came from. Key figure: 14 million records linked in one graph, each link traced to its source (36 million links between contracts, companies, people and public bodies). Selected work: One procurement platform fed by notices, EU tenders, EU law, the official gazette and complaints-council rulings; Company-registry and officials’ declaration records joined to procurement data in one graph. - **Ledgers** (2018 → 2023): I built settlement and custody back-ends for banks and digital-asset platforms in Switzerland and Canada, and lending and staking contracts for blockchain products. These systems hold other people’s money, so a bug is a direct loss. Key figure: $0 lost from the contracts and platforms I built (2018–2023 · architect, then VP of Engineering · Go and Solidity · references on request). Selected work: Lending and staking contracts, a blockchain node and a bridge between chains; Security audits of other teams’ contracts; Research on detecting and blocking front-running (reordering transactions to take value). - **Back-ends** (2014 → now): I build the APIs, payment services, queues and databases that products run on, in Go and Java, with automated builds and monitoring. Key figure: 100,000 users at peak on a launch platform I built in 3–4 months (2018–2023 · Go back-end · references on request). Selected work: A file catalogue that indexed 8.2 million files across 20 drives (ONCE, my own product); A Java payments service on Stripe; A gateway that limits how much each API key can call and spend on AI models. - **Devices** (2016 → 2018): Firmware for a connected home-security camera: encrypted over-the-air updates, a watchdog that restarts it after a fault, live video, and face recognition that runs on the camera itself. Key figure: 2 cores on the camera’s processor ran face recognition, tuned to each household from a short video (2016–2018 · a crowdfunded camera, taken through a full production run). Selected work: Live video over WebRTC, Bluetooth (BLE) links, and MQTT messaging between camera and cloud. ## Client systems (a sample of four; names withheld, sector and problem only) - **Bid documents drafted by agents, approved by people** (Public procurement · SaaS; client system · in production): Companies bidding for public contracts spend days assembling each tender pack from long, inconsistent documentation, and one missed requirement can disqualify a bid. Built: It reads the official sources, matches each new tender to the companies it suits and drafts the pack in stages: one agent plans it, writers draft the sections in parallel, and an independent reviewer sends weak work back for repair. Controls: The system only drafts; a person reviews and submits every document. In a test on real tenders with 17 planted defects, the reviewer agent caught all 17. Evidence: 8 kinds of official source, with new procurement notices read every 30 seconds. In production with hundreds of users, and it has passed a legal review. Role: Architecture, back-end and agent design. - **Multilingual voice and chat agents, live with customers** (Sales, marketing and legal services; client system · in production): Many customers would rather talk than fill in a form. A voice agent that lags, mishears accents, talks over people or makes up answers is worse than the form, and one left unwatched can run up costs quickly. Built: Each agent listens and replies in one continuous stream, waits until the caller has finished speaking, and can call tools mid-conversation. The sales agents hand a conversation to a person when it needs one. Controls: Answers draw on the client’s documents and, in some deployments, the web, and each one is marked with its source. The deployments limit how many sessions a visitor can open and how many run at once, and have a daily budget, a kill switch and backup models. The agents are tested against fixed sets of conversations. Evidence: 250,000+ minutes of live sessions since May 2026, and more than 40,000 sessions in the 60 days to mid-November. Live since May 2026 in English, Romanian, French and many other widely spoken languages. Monthly use has grown from about 7,000 minutes in May to more than 60,000 in the first half of November. Role: Architecture and build. - **Camera analytics that never store video** (Live events and industrial sites · Safety; client system · from on-site pilot to production): Event organisers need crowd density and flow in real time. Safety teams need fast alerts when someone is missing protective equipment or goes down. Neither should create a surveillance archive. Built: Live streams are analysed in memory on a GPU: crowd counts, flows and alerts for events, and protective-equipment and person-down alerts for industrial sites, shown on a bilingual operator console. Controls: No video is recorded. An audit command, run in the client’s deployment, fails if it finds a video file, an unblurred image or an image with no database record behind it. Where evidence is needed, only blurred stills are stored; every access to them is logged, and they are deleted after a fixed period. Evidence: 0 video recordings kept, checked by an audit command in the client’s deployment. Piloted on site at a large music festival in Romania with the organisers and public authorities, and now going into production for an industrial client. One machine analyses 16 live camera streams. Role: Architecture and build. - **Romania’s public records, linked and traced to source** (Public records · Romania; part client work, part own research): Procurement, company-registry and officials’ declaration records are public, but split across systems and too large to search by hand. An investigation needs every link traced to its source. Built: One graph joined from the three sources, every link carrying its source. An agent with 12 tools answers questions by searching the graph, building views on demand and pulling out the records it needs. The agent and the extraction models run on premises, on a GPU server. Controls: The agent reaches the data only through a gateway that decides which queries it may run. Evidence: 14 million records in one graph, with 36 million links between contracts, companies, people and public bodies. Runs on Romanian procurement, company-registry and declaration data: 6.4 million contracts and 4.1 million companies so far, imported at more than 10,000 rows a second. Every change to the agent must pass a fixed set of 16 cases before it ships. Role: Architecture and build. ## Working together - **Who you work with:** Me, directly. On larger builds, engineers from my own network, always named in the contract. - **Where:** Remote from Cluj-Napoca, on Eastern European Time. On site for kick-off, key reviews and handover. - **Pricing:** There is no price list; the scope sets the cost. A fixed price or a monthly fee, agreed in writing. A sprint can turn into a build, and a build into a long-term engagement. - **Your data:** You decide where it goes, in writing: EU-only providers, the major US model providers under business terms, or models on your own hardware. I build the controls that GDPR and the EU AI Act expect, such as human oversight, logging, data minimisation and fixed retention, and document them for your compliance review. - **References:** Available after a first conversation, with the client’s consent. ## Own products - **Vigildot** (Beta · Private): Monitoring that your coding agent sets up for you. Self-hosted monitoring for uptime, errors, servers and AI spend. A coding agent can install it, wire it into an app and show that one real operation is tracked from start to finish. Alerts go to the on-call person’s phone, with one-tap acknowledgement. - **ONCE** (Beta · Pre-release): Every file. Exactly once. A catalogue of every drive you own. ONCE finds duplicate files, tells you which copies are safe to delete because a verified copy exists elsewhere, and keeps one archive. Built for photographers and videographers first. On a set of 20 drives it found that 52% of 30.6 TB was duplicate copies. - **Agent API fleet** (Prototype · Private): Checks and receipts for machine-to-machine AI. 26 small services built for AI agents to call and pay for one request at a time: checks that a service behaves as its listing says, safety checks that say no when unsure, and receipts anyone can verify. - **Teal** (Alpha · Private): A voice assistant that sends agents on errands. Talk to it as you would on a phone call. It sends agents on your errands and interrupts itself to tell you what they found. For iOS and Android. - **Provinciae** (Live · Public): A live world that people and AI agents play together. A real-time multiplayer simulation in the browser. AI agents join through MCP, the standard way agents connect to outside tools, and play alongside people. It is live, with real players in the world today. In a 24-hour load test with 30 builders working in parallel, 99% of world updates finished in under 22 ms. https://provinciae.online ## History - **2014 to 2016 · Engineering:** I started as a back-end engineer, building data systems in Java and C++. - **2016 to 2018 · Devices & IoT:** Firmware for a crowdfunded home-security camera, from prototype through a full production run: encrypted updates, live video, and face recognition running on the camera itself. - **2018 to 2023 · Leadership & fintech:** I founded Redcom Software in 2018. Since then I have been CTO twice, VP of Engineering (2021–2023) and technical architect four times, and co-founded two companies. Most of that work was settlement, custody and ledger back-ends for banks and digital-asset platforms in Switzerland and Canada, where no funds were ever lost on a system I built. - **2023 to now · AI for clients:** Voice and chat agents with more than 250,000 live minutes, bid-drafting agents used by hundreds of people, a public-records graph of 14 million records and camera analytics that keep no video, for clients in Europe and the Middle East. Alongside them, my own products and research. ## Lab notes Research run alongside client work, newest first. Each note gives the method and the numbers, including results that went against the author. - [When Fable reviews, Qwen fixes more](https://redcomsoftware.com/notes/advisor-review/) (November 2026 · Controlled experiment): On 80 real issues from open-source projects, a published 4-bit fine-tune of Qwen3.8 27B, served on my own hardware, resolved 65 of the 78 that ran cleanly. When Claude Fable 5.2 reviewed each first patch and sent back the ones it doubted, Qwen resolved 74, fixing 9 and breaking none. A follow-up found that Sonnet 5.5 briefs did about as well as Fable’s at 18% of the cost. The gain needs a second run, and each run had twice the benchmark’s time limit, and Fable may have seen these public fixes in training. - [A local model counted the corrections. A hand count disagreed.](https://redcomsoftware.com/notes/agent-transcripts/) (November 2026 · Measurement): A 27B model on the author’s own DGX Spark, behind masking and a sandbox, read 20,174 prompts from fourteen months of Claude Code history on one of the author’s machines overnight and counted one request in eight as a correction. A blind hand count by the author on a small 2026 sample found nearly one in four, and the model missed most of those; Claude Opus 5.5 did little better. The hand check stopped a number that may be too low from being reported unchecked. The hand count hints at a fall late in 2026 and the model’s rate edges down, though the samples are too small to be sure. - [When in doubt, the agent doesn’t pay](https://redcomsoftware.com/notes/fail-closed/) (August 2026 · Build note · test build): Before an AI agent pays for an API call over x402 (pay-per-request inside HTTP), a gate I built checks the live price and payee against the service’s listing and the buyer’s rules, and never takes missing evidence as a yes. In fixture tests on 40 synthetic providers, 35 built to misbehave, every one whose live terms drifted from its listing was denied and all five clean controls were allowed; nothing is live yet. - [Sharper on one dataset, reversed on the next](https://redcomsoftware.com/notes/frozen-model-depths/) (July 2026 · Stopped · negative result): A side network 0.2% the size of a frozen language model read four of its layers and made its answers more decisive than a same-size control did, at all three fresh seeds tested. On a second dataset the effect reversed and no change to generated text could be shown, so I stopped under a rule written before the deciding runs. ## Company - REDCOM SOFTWARE SRL · CUI 38869281 · VAT RO38869281 · J12/549/2018 · founded 2018 - An EU company, registered for VAT. Invoicing in EUR or USD.