Dr. Tali Režun, a non-developer entrepreneur and academic specializing in agentic systems, has published a detailed case study describing how he delivered a working minimum viable product for a real-time voice-driven video avatar platform in eight weeks. Working alone, he directed more than 400 AI agents through roughly 300 orchestrator sessions that produced 367 commits. The resulting system, called ARIA (AI Retention and Interaction Assistant), comprises three applications and a shared package in a single monorepo totaling a little over 17,000 lines of code, plus around 14 technical documents prepared for the client’s developers.
The client, whose company employs more than 500 people including a large technical staff, had received an internal engineering estimate of six senior developers working eight to twelve months for the same MVP. Režun completed the build without writing the application code himself. The agents handled implementation while he focused on architecture decisions, context management, and verification.
ARIA is a live, talking, animated companion that embeds inside an existing digital product. Users speak to it and receive natural-voice replies from a lip-synced video face at conversational latency. The system consumes a live event stream from the host platform, maintains memory across sessions, and supports frictionless embedding via snippet, iframe, or hosted URL. Measured end-to-end performance once a session is live shows the first spoken response after the user stops talking arriving in approximately 480 milliseconds.
Technical Architecture and Orchestration
The build relied on Intent, an orchestration environment from Augment Code. A coordinator agent reads the codebase and documentation, maintains a living specification, and decomposes work into discrete tasks. Implementor agents execute those tasks in isolated git worktrees to avoid file collisions. A verifier checks results against the specification before human review. Models were assigned by role rather than treated as interchangeable: frontier models handled coordination and complex reasoning, while other models performed implementation and independent cross-vendor auditing.
Documentation served as the control plane. Every non-trivial change began as a written brief specifying affected files, invariants, and forbidden alterations. Architecture documents contained explicit contributor guardrails that agents were required to follow. Context windows of the primary models (approximately 200,000 tokens) required frequent migration. At roughly 80–90 percent utilization, the outgoing orchestrator produced a structured handoff containing decisions with rationale, known traps, open work, and role instructions. The incoming session received only the documents relevant to the next tasks. These migrations occurred about 300 times over the project.
Core conversation-flow logic separated silent context updates from spoken triggers, applied priority tiers to events, queued and deduplicated speech requests, imposed rate limits and grace periods, and managed memory across three layers: cross-session summaries, in-session checkpoints, and reconnect continuity. Human judgment defined these rules; agents then implemented the resulting specifications.
Empirical Outcomes and Costs
The delivered MVP runs on a single DigitalOcean droplet in an EU region behind Nginx, with PM2 managing Node processes. The same codebase supports local development, self-hosted tenant deployment, or multi-tenant SaaS through configuration. Scaling analysis indicated that external voice and avatar provider concurrency limits, not the application infrastructure, form the primary ceiling.
API costs for Anthropic and OpenAI models peaked near $300 per day during intensive parallel phases. Even assuming peak spend every build day, total credit expenditure remained in the range of $12,000 to $13,000. Režun notes that this figure excludes his own intensive full-time effort, external service fees, and the two years of prior experimentation required to develop the orchestration method. Traditional staffing of six senior developers for eight to twelve months would represent a substantially larger six-figure outlay before the idea had been validated.
Handover documentation was produced continuously rather than as a final afterthought, because the same documents guided the agents throughout construction. The client’s existing developers now maintain and extend the system using that material.
Industry Context and Remaining Challenges
Režun emphasizes that the primary constraint was not model capability but context engineering and coordination. Parallel agents without strong isolation and shared, up-to-date specifications produce collisions and inconsistent assumptions. Quiet agent failures—locally plausible code that fails only under combined conditions—required human symptom description and multi-file audits. Context migration itself consumed significant energy and did not scale beyond a single highly attentive operator.
He later developed an open-source tool, The Curator, to address the memory and handoff burden by providing persistent, MCP-accessible project state that agents can read and update. The case study illustrates a broader shift in which experienced operators can test software ideas at speeds and costs previously reserved for validated production initiatives, while human judgment remains essential for architectural decisions, conversational design, and quality verification.
Sources: Dr. Tali Režun’s case study on Substack; related post by Dr. Tali Režun on X.
